Numerical Approximation Methods For Elliptic Boundary Value Problems PDF

Numerical Approximation Methods
for Elliptic Boundary Value Problems

Olaf Steinbach
Numerical Approximation
Methods for Elliptic Boundary
Value Problems
Finite and Boundary Elements

Olaf Steinbach
Institute of Computational Mathematics
Graz University of Technology
Austria
Originally published in the German language by B.G. Teubner Verlag as “Olaf Steinbach:
Numerische Näherungsverfahren für elliptische Randwertprobleme. 1. Auflage (1st ed.)”.

c B.G. Teubner Verlag|GWV Fachverlage GmbH, Wiesbaden 2003
English version published by Springer Science+Business Media, LLC
ISBN 978-0-387-31312-2 e-ISBN 978-0-387-68805-3
Library of Congress Control Number: 2007936614
Mathematics Subject Classification (2000): 65N30, 65N38
c 2008 Springer Science+Business Media, LLC

All rights reserved. This work may not be translated or copied in whole or in part without the written
permission of the publisher (Springer Science+Business Media, LLC, 233 Spring Street, New York, NY
10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection
with any form of information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed is forbidden.
The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are
not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject
to proprietary rights.
Printed on acid-free paper.
9 8 7 6 5 4 3 2 1
springer.com
Preface
Finite and boundary element methods belong to the most used numerical dis-
cretization methods for the approximate solution of elliptic boundary value
problems. Finite element methods (FEM) are based on a variational formu-
lation of the partial differential equation to be solved. The definition of a
conforming finite dimensional trial space requires an appropriate decomposi-
tion of the computational domain into finite elements. The advantage of using
finite element methods is their almost universal applicability, e.g. when con-
sidering nonlinear partial differential equations. Contrary, the use of boundary
element methods (BEM) requires the explicit knowledge of a fundamental so-
lution, which allows the transformation of the partial differential equation to a
boundary integral equation to be solved. The approximate solution then only
requires a decomposition of the boundary into boundary elements. Bound-
ary element methods are often used to solve partial differential equations
with (piecewise) constant coefficients, and to find solutions of boundary value
problems in exterior unbounded domains. In addition, direct boundary el-
ement methods provide a direct computation of the complete Cauchy data
which are the real target functions in many applications. In finite element
methods, the Cauchy data can be computed by using Lagrange multipliers
and by solving related saddle point problems. By combining both discretiza-
tion methods it is possible to profit from the advantages of both methods.
Although the aim of this book is to give a unified introduction into fi-
nite and boundary element methods, the main focus of the presentation is on
the numerical analysis of boundary integral equation methods. Therefore, we
only consider some linear model problems such as the potential equation, the
system of linear elasticity, the Stokes system, and the Helmholtz equation.
When considering the above mentioned elliptic boundary value problems it is
possible to describe and to analyze finite and boundary element methods in
a unified manner. After the description of the model problems, we introduce
the function spaces which are needed later. Then we discuss variational meth-
ods for the solution of operator equations with and without side conditions. In
particular, this also includes the formulation of saddle point problems by using
VI Preface
Lagrange multipliers. The variational formulation of boundary value problems

is the basis of finite element methods, but on the other hand, domain varia-
tional methods are also needed in the analysis of boundary integral operators.
After the computation of fundamental solutions we define certain boundary
integral operators, and analyze their mapping properties such as bounded-
ness and ellipticity. For the solution of different boundary value problems we
then describe and analyze different boundary integral equations to find the
complete Cauchy data. Numerical discretization methods are first formulated
and investigated in a more abstract setting. Afterwards, appropriate finite
dimensional trial spaces are constructed, and corresponding approximation
properties are given. For the solution of mixed boundary value problems we
then discuss different finite and boundary element methods. In particular, we
investigate the properties of the associated linear systems of algebraic equa-
tions. For this, we also describe appropriate preconditioned iterative solution
strategies where the proposed preconditioning techniques involves both the
preconditioning with integral operators of the opposite order, and a hierarchi-
cal multilevel preconditioner. Since the Galerkin discretization of boundary
integral operators leads to dense stiffness matrices, fast boundary element
methods are used to obtain an almost optimal complexity of storage and of
matrix by vector multiplications. Finally, we describe domain decomposition
methods to handle partial differential equations with jumping coefficients, and
to couple and parallelize different discretization techniques such as finite and
boundary element methods.
Boundary element methods are a well established numerical method for
elliptic boundary value problems as discussed in this textbook. For the sake
of simplicity in the presentation, we only consider the case of linear and self–
adjoint partial differential equations. For more general partial differential op-
erators one has to consider the fundamental solution of the formally adjoint
operator. While the existence of fundamental solutions can be ensured for a
large class of partial differential operators their explicit knowledge is manda-
tory for a numerical realization. For nonlinear partial differential equations
there also exist different approaches to formulate boundary integral equation
methods which are often based on the use of volume potentials to cover the
nonlinear terms.
While there exists a rather huge number of textbooks on finite element
methods, e.g. [5, 21, 31, 41, 57] just to mention a few of them, much less
is available on boundary integral and boundary element methods. For the
analysis of boundary integral operators related to elliptic partial differential
equations we refer to [3, 60, 81, 88, 102, 103] while for the numerical analysis
of boundary element methods we mention [39, 124, 126, 135]. In addition, the
references [9, 15, 19, 30, 61, 77] are on practical aspects of the use of boundary
element methods in engineering. In [74, 98, 125, 158] one may find recent
results on the use of advanced boundary element algorithms. For a detailed
description of fast boundary element methods starting from the basic ideas
Preface VII
and proceeding to their practical realization see [117] where also numerous
examples are given.
Since the aim of this textbook is to give a unified introduction into finite
and boundary element methods, not all topics of interest can be discussed.
For a further reading we refer, e.g., to [51, 130] for hp finite element methods,
to [2, 8, 10, 153] for a posteriori error estimators and adaptive finite element
methods, and to [24, 69] for multigrid methods. In the case of boundary el-
ement methods we refer, e.g., to [17, 99, 100, 131, 146] for hp methods, to
[37, 38, 56, 128, 129] for a posteriori error estimators and adaptive methods,
and to [96, 111, 147] for multigrid and multilevel methods. Within this text-
book we also do not discuss the matter of numerical integration, for this we
refer to [55, 67, 87, 93, 115, 123, 132, 148] and the references given therein.
This textbook is based on lectures on the numerical solution of elliptic
partial differential equations which I taught at the University of Stuttgart, at
the Technical University of Chemnitz, at the Johannes Kepler University of
Linz, and at Graz University of Technology. Chapters 1–4, 8, 9, 11, and 13 can
be used for an introductory lecture on finite element methods, while chapters
1–8, 10, 12, and 13 are on the basics of the boundary element method. Chapter
14 gives an overview on fast boundary element methods. Besides the use as
a complementary textbook it is also recommended for self–study for students
and researchers, both in applied mathematics, in scientific computing, and in
computational engineering.
It is my great pleasure to thank W. L. Wendland for his encouragement
and support over the years. Many results of our joint work influenced this
book. Special thanks go to J. Breuer and G. Of, who read the original German
manuscript and made valuable comments and corrections. This text book was
originally published in a German edition [140]. Once again I would like to
thank J. Weiss and B. G. Teubner for the fruitful cooperation.
When preparing the English translation I got many responses, suggestions
and hints on the German edition. I would like to thank all who helped to
improve the book. In particular I thank G. Of, S. Engleder and D. Copeland
who read the English manuscript. Finally I thank Springer New York for the
cooperation and the patience when preparing this book.
Graz, August 2007 Olaf Steinbach

Contents
1 Boundary Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.1 Potential Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Linear Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Plane Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.2.2 Incompressible Elasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.3 Stokes System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.4 Helmholtz Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
1.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.1 The Spaces C k (Ω), C k,κ (Ω) and Lp (Ω) . . . . . . . . . . . . . . . . . . . . 19
2.2 Generalized Derivatives and Sobolev Spaces . . . . . . . . . . . . . . . . . 22
2.3 Properties of Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.4 Distributions and Sobolev Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . 29
2.5 Sobolev Spaces on Manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3 Variational Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.1 Operator Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.2 Elliptic Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.3 Operators and Stability Conditions . . . . . . . . . . . . . . . . . . . . . . . . 48
3.4 Operator Equations with Constraints . . . . . . . . . . . . . . . . . . . . . . 50
3.5 Mixed Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6 Coercive Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
4 Variational Formulations of Boundary Value Problems . . . . . 59

4.1 Potential Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
4.1.1 Dirichlet Boundary Value Problem . . . . . . . . . . . . . . . . . . 61
4.1.2 Lagrange Multiplier Methods . . . . . . . . . . . . . . . . . . . . . . . 64
4.1.3 Neumann Boundary Value Problem . . . . . . . . . . . . . . . . . . 67
4.1.4 Mixed Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . 70
X Contents
4.1.5 Robin Boundary Value Problems . . . . . . . . . . . . . . . . . . . . 71

4.2.1 Dirichlet Boundary Value Problem . . . . . . . . . . . . . . . . . . 76
4.2.2 Neumann Boundary Value Problem . . . . . . . . . . . . . . . . . . 77
4.2.3 Mixed Boundary Value Problems . . . . . . . . . . . . . . . . . . . . 79
4.3 Stokes Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
5 Fundamental Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
5.1 Laplace Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
5.3 Stokes Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
5.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6 Boundary Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

6.1 Newton Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
6.2 Single Layer Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.3 Adjoint Double Layer Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
6.4 Double Layer Potential . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.5 Hypersingular Boundary Integral Operator . . . . . . . . . . . . . . . . . 128
6.6 Properties of Boundary Integral Operators . . . . . . . . . . . . . . . . . . 136
6.6.1 Ellipticity of the Single Layer Potential . . . . . . . . . . . . . . 139
6.6.2 Ellipticity of the Hypersingular Boundary Integral
Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
6.6.3 Steklov–Poincaré Operator . . . . . . . . . . . . . . . . . . . . . . . . . 148
6.6.4 Contraction Estimates of the Double Layer Potential . . . 149
6.6.5 Mapping Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152
6.8 Stokes System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
6.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169
7 Boundary Integral Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

7.1 Dirichlet Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . . . . . 172
7.2 Neumann Boundary Value Problem . . . . . . . . . . . . . . . . . . . . . . . . 175
7.3 Mixed Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
7.4 Robin Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
7.5 Exterior Boundary Value Problems . . . . . . . . . . . . . . . . . . . . . . . . 181
7.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
Contents XI
8 Approximation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

8.1 Galerkin–Bubnov Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
8.2 Approximation of the Linear Form . . . . . . . . . . . . . . . . . . . . . . . . . 190
8.3 Approximation of the Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . 191
8.4 Galerkin–Petrov Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
8.5 Mixed Formulations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
8.6 Coercive Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
8.7 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
9 Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

9.1 Reference Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203
9.2 Form Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211
9.3 Trial Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
9.4 Quasi Interpolation Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
9.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
10 Boundary Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

10.1 Reference Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
10.2 Trial Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
11 Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

11.3 Finite Element Methods with Lagrange Multipliers . . . . . . . . . . 255
11.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261
12 Boundary Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

12.3 Mixed Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
12.4 Robin Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 287
12.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
13 Iterative Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

13.1 The Method of Conjugate Gradients . . . . . . . . . . . . . . . . . . . . . . . 291
13.2 A General Preconditioning Strategy . . . . . . . . . . . . . . . . . . . . . . . 299
13.2.1 An Application in Boundary Element Methods . . . . . . . . 302
13.2.2 A Multilevel Preconditioner
in Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . 306
13.3 Solution Methods for Saddle Point Problems . . . . . . . . . . . . . . . . 319
14 Fast Boundary Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 327

14.1 Hierarchical Cluster Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
14.2 Approximation of the Stiffness Matrix . . . . . . . . . . . . . . . . . . . . . . 332
14.2.1 Taylor Series Representations . . . . . . . . . . . . . . . . . . . . . . . 336
14.2.2 Series Representations of the Fundamental Solution . . . . 340
14.2.3 Adaptive Cross Approximation . . . . . . . . . . . . . . . . . . . . . . 344
XII Contents
14.3 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

14.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 366
15 Domain Decomposition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 367
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
1
Boundary Value Problems
In this chapter we describe some stationary boundary value problems with

self–adjoint partial differential operators of second order. As simple model
problems we consider the scalar potential equation and the Helmholtz equa-
tion, while for a system we consider the equations of linear elasticity and, as
for incompressible materials, the Stokes system.
1.1 Potential Equation

Let Ω ⊂ Rd (d = 2, 3) be a bounded and simply connected domain with
sufficiently smooth boundary Γ = ∂Ω, and n(x) is the exterior unit normal
vector which is defined almost everywhere for x ∈ Γ . For x ∈ Ω we consider a
self–adjoint linear partial differential operator of second order which is applied
to a scalar real valued function u,
d
∂ ∂
(Lu)(x) := − aji (x) u(x) + a0 (x)u(x) . (1.1)
i,j=1
∂xj ∂xi
The coefficient functions aji (x) are assumed to be sufficient smooth satisfying
aij (x) = aji (x) for all i, j = 1, . . . , d, x ∈ Ω. Partial differential operators
of the form (1.1) are used to model, for example, the static heat transfer,
electrostatic potentials, or ideal fluids.
For a classification [79] of scalar partial differential operators L we consider
the real eigenvalues λk (x) of the symmetric coefficient matrix
d
A(x) = (aij (x))i,j=1 , x ∈ Ω.
The partial differential operator L is called elliptic at x ∈ Ω iff λk (x) > 0 is
satisfied for all k = 1, . . . , d. If this condition is satisfied for all x ∈ Ω, then L
is elliptic in Ω. If there exists a uniform lower bound λ0 > 0 satisfying
λk (x) ≥ λ0 for k = 1, . . . , d and for all x ∈ Ω, (1.2)
the partial differential operator L is called uniformly elliptic in Ω.
2 1 Boundary Value Problems
The starting point for what follows is the well known theorem of Gauss
and Ostrogradski, i.e.

∂
f (x)dx = γ0int f (x)ni (x)dsx , i = 1, . . . , d,
∂xi
Ω Γ
where
γ0int f (x) := lim f (
x) for x ∈ Γ = ∂Ω (1.3)
x→x∈Γ
Ω
is the interior boundary trace of a given function f (x), x ∈ Ω.

For sufficiently smooth functions u, v we consider f (x) = u(x)v(x) to
obtain the formula of integration by parts,

∂ ∂
v(x) u(x)dx = γ0int u(x)γ0int v(x)ni (x)dsx − u(x) v(x)dx. (1.4)
∂xi ∂xi
Ω Γ Ω
When multiplying the partial differential operator (1.1) with a sufficiently

smooth test function v, and integrating the result over Ω, this gives
d

∂ ∂
(Lu)(x)v(x)dx = − aji (x) u(x) v(x)dx .
i,j=1 Ω
∂xj ∂xi
Ω
Applying integration by parts, see (1.4), we obtain

d
∂ ∂
(Lu)(x)v(x)dx = aji (x) u(x) v(x)dx
i,j=1 Ω
∂xi ∂xj
Ω
d

∂
− nj (x)γ0int aji (x) u(x) γ0int v(x)dsx ,
i,j=1 Γ
∂xi
and therefore Green’s first formula

a(u, v) = (Lu)(x)v(x)dx + γ1int u(x)γ0int v(x)dsx (1.5)
Ω Γ
by using the symmetric bilinear form

d
∂ ∂
a(u, v) := aji (x) u(x) v(x)dx (1.6)
i,j=1 Ω
∂xi ∂xj
as well as the interior conormal derivative

⎡ ⎤
d
∂
γ1int u(x) := lim ⎣ nj (x)aji (
x) x)⎦
u( for x ∈ Γ. (1.7)
Ωx→x∈Γ
i,j=1
∂
x i
1.1 Potential Equation 3
As in (1.5) we find, by exchanging the role of u and v, the analogue Green’s

formula

a(u, v) = a(v, u) = (Lv)(x)u(x)dx + γ1int v(x)γ0int u(x)dsx .
Ω Γ
Combining this with (1.5) we therefore obtain Green’s second formula

(Lu)(x)v(x)dx + γ1int u(x)γ0int v(x)dsx (1.8)
Ω Γ

= (Lv)(x)u(x)dx + γ1int v(x)γ0int u(x)dsx
Ω Γ
which holds for arbitrary but sufficiently smooth functions u and v.
Example 1.1. For the special choice aij (x) = δij where δij is the Kronecker
delta with δij = 1 for i = j and δij = 0 for i = j, the partial differential
operator (1.1) is the Laplace operator
d
∂2
(Lu)(x) = −∆u(x) := − u(x) for x ∈ Rd . (1.9)
i=1
∂x2i
The associated conormal derivative (1.7) coincides with the normal derivative
∂
γ1int u(x) = u(x) := n(x) · ∇u(x) for x ∈ Γ.
∂nx
Let Γ = Γ D ∪ Γ N ∪ Γ R be a disjoint decomposition of the boundary

Γ = ∂Ω. The boundary value problem is to find a scalar function satisfying
the partial differential equation
(Lu)(x) = f (x) for x ∈ Ω, (1.10)
the Dirichlet boundary conditions
γ0int u(x) = gD (x) for x ∈ ΓD , (1.11)
the Neumann boundary conditions
γ1int u(x) = gN (x) for x ∈ ΓN , (1.12)
and the Robin boundary conditions
γ1int u(x) + κ(x)γ0int u(x) = gR (x) for x ∈ ΓR (1.13)
where f , gD , gN , gR , and κ are some given functions. The boundary value

problem (1.10) and (1.11) with Γ = ΓD is called a Dirichlet boundary value
problem, while the boundary value problem (1.10) and (1.12) with Γ = ΓN is
called a Neumann boundary value problem. In the case Γ = ΓR the problem
(1.10) and (1.13) is said to be a Robin boundary value problem. In all other
cases we have to solve boundary value problems with boundary conditions of
mixed type. Note that one may also consider nonlinear Robin type boundary
conditions [119]
γ1int u(x) + G(γ0int u, x) = gR (x) for x ∈ ΓR
where G(v, ·) is some given function, e.g. G(v, ·) = v 3 or G(v, ·) = v 4 .

In the classical approach, the solution of the boundary value problem
(1.10)–(1.13) has to be sufficiently differentiable, in particular we require
u ∈ C 2 (Ω) ∩ C 1 (Ω ∪ ΓN ∪ ΓR ) ∩ C(Ω ∪ ΓD )
where we have to assume that the given data are sufficiently smooth. For
results on the unique solvability of boundary value problems in the classical
sense we refer, for example, to [92].
In the case of a Neumann boundary value problem (1.10) and (1.12) addi-
tional considerations are needed to investigate the solvability of the boundary
value problem. Obviously, v1 (x) = 1 for x ∈ Ω is a solution of the homoge-
neous Neumann boundary value problem
(Lv1 )(x) = 0 for x ∈ Ω, γ1int v1 (x) = 0 for x ∈ Γ. (1.14)
Applying Green’s second formula (1.8) we then obtain the orthogonality

(Lu)(x)dx + γ1int u(x)dsx = 0 . (1.15)
Ω Γ
When considering the Neumann boundary value problem (1.10) and (1.12),
(Lu)(x) = f (x) for x ∈ Ω, γ1int u(x) = gN (x) for x ∈ Γ, (1.16)
and using the orthogonality (1.15) for the given data f and gN , we have to
assume the solvability condition

f (x)dx + gN (x)dsx = 0. (1.17)
Ω Γ
Since there exists a non–trivial solution v1 (x) = 1 for x ∈ Ω of the homoge-

neous Neumann boundary value problem (1.14), we conclude that the solution
of the Neumann boundary value problem (1.16) is only unique up to an ad-
ditive constant. Let u be a solution of (1.16). Then, for any α ∈ R we can
define
(x) = u(x) + α for x ∈ Ω
u
1.2 Linear Elasticity 5
to be also a solution of the Neumann boundary value problem (1.16). The

constant α ∈ R is uniquely determined when requiring an additional scaling
condition on the solution u of (1.16), e.g.

u(x) dx = 0, or γ0int u(x) dsx = 0.
Ω Γ
1.2 Linear Elasticity

As an example for a system of partial differential equations we consider the
system of linear elasticity. For any x ∈ Ω we have to find a vector valued
function u(x) with components ui (x), i = 1, 2, 3, describing the displacements
of an elastic body where we assume a reversible, isotropic and homogeneous
material behavior. In particular, we consider the equilibrium equations
3
∂
− σij (u, x) = fi (x) for x ∈ Ω, i = 1, 2, 3. (1.18)
j=1
∂xj
In (1.18), σij (u, x) are the components of the stress tensor which is linked to
the strain tensor eij (u, x) by Hooke’s law,
3
Eν E
σij (u, x) = δij ekk (u, x) + eij (u, x) (1.19)
(1 + ν)(1 − 2ν) 1+ν
k=1
for x ∈ Ω, i, j = 1, 2, 3, and with the Young modulus E > 0 and with the
Poisson ratio ν ∈ (0, 12 ). Moreover, when assuming small deformations the
linearized strain tensor is given by

1 ∂ ∂
eij (u, x) = uj (x) + ui (x) for x ∈ Ω, i, j = 1, 2, 3. (1.20)
2 ∂xi ∂xj
Multiplying the equilibrium equations (1.18) with a test function vi , integrat-
ing over Ω, and applying integration by parts, this gives for i = 1, 2, 3
3
∂
fi (x)vi (x)dx = − σij (u, x)vi (x)dx
j=1
∂xj
Ω Ω

3
3
∂
= σij (u, x) vi (x)dx − nj (x)σij (u, x)vi (x)dsx .
∂xj
Ω j=1 Γ j=1
Taking the sum for i = 1, 2, 3 we obtain Betti’s first formula

3
∂
− σij (u, x)vi (x)dx = a(u, v) − γ0int v(x) γ1int u(x)dsx (1.21)
i,j=1
∂x j
Ω Γ
with the bilinear form

3
∂
a(u, v) := σij (u, x) vi (x)dx
∂xj
Ω i,j=1

3
1 ∂ ∂
= σij (u, x) vi (x) + vj (x) dx
2 ∂xj ∂xi
Ω i,j=1

3
= σij (u, x)eij (v, x)dx, (1.22)
Ω i,j=1
and with the conormal derivative

3

(γ1int u)i (x) := σij (u, x)nj (x) for x ∈ Γ, i = 1, 2, 3. (1.23)
j=1
Inserting the strain tensor (1.20) as well as Hooke’s law (1.19) into the equi-
librium equations (1.18) we obtain a system of partial differential equations
where the unknown function is the displacement field u. First we have
3
3
∂
ekk (u, x) = uk (x) =: div u(x)
∂xk
k=1 k=1
and therefore
Eν E ∂
σii (u, x) = div u(x) + ui (x),
(1 + ν)(1 − 2ν) 1 + ν ∂xi

E ∂ ∂
σij (u, x) = uj (x) + ui (x) for i = j.
2(1 + ν) ∂xi ∂xj
From this we obtain

E Eν E ∂
− ∆ui (x) − + div u(x) = fi (x)
2(1 + ν) (1 + ν)(1 − 2ν) 2(1 + ν) ∂xi
for x ∈ Ω, i = 1, 2, 3. By introducing the Lamé constants

Eν E
λ = , µ = (1.24)
(1 + ν)(1 − 2ν) 2(1 + ν)
we finally conclude the Navier system
−µ∆u(x) − (λ + µ)grad div u(x) = f (x) for x ∈ Ω. (1.25)

The bilinear form (1.22) can be written as

3

a(u, v) = σij (u, x)eij (v, x)dx
i,j=1 Ω

3
= 2µ eij (u, x)eij (v, x)dx + λ div u(x) div v(x) dx (1.26)
Ω i,j=1 Ω
implying the symmetry of the bilinear form a(·, ·).

The first component of the conormal derivative (1.23) is
3

(γ1int u)1 (x) = σ1j (u, x)nj (x)
j=1

∂ ∂ ∂
= λ div u(x) + 2µ u1 (x) n1 (x) + µ u2 (x) + u1 (x) n2 (x)
∂x1 ∂x1 ∂x2

∂ ∂
+µ u3 (x) + u1 (x) n3 (x)
∂x1 ∂x3

∂ ∂ ∂
= λ div u(x) n1 (x) + 2µ u1 (x) + µ u2 (x) − u1 (x) n2 (x)
∂nx ∂x1 ∂x2

∂ ∂
+µ u3 (x) − u1 (x) n3 (x).
∂x1 ∂x3
From this we obtain the following representation of the boundary stress op-
erator for x ∈ Γ ,
∂
γ1int u(x) = λ div u(x) n(x) + 2µ u(x) + µ n(x) × curl u(x). (1.27)
∂nx
In many applications of solid mechanics the boundary conditions (1.11) and
(1.12) are given within their components, i.e.
γ0int ui (x) = gD,i (x) for x ∈ ΓD,i ,

(1.28)
(γ1int u)i (x) = gN,i (x) for x ∈ ΓN,i ,
where Γ = Γ D,i ∪ Γ N,i for i = 1, 2, 3.

The non–trivial solutions of the homogeneous Neumann boundary value
problem
−µ∆u(x) − (λ + µ)grad div u(x) = 0 for x ∈ Ω, γ1int u(x) = 0 for x ∈ Γ
are given by the rigid body motions v k ∈ R where

⎧⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞ ⎛ ⎞⎫
⎨ 1 0 0 −x2 0 x3 ⎬
R = span ⎝ 0 ⎠ , ⎝ 1 ⎠ , ⎝ 0 ⎠ , ⎝ x1 ⎠ , ⎝ −x3 ⎠ , ⎝ 0 ⎠ . (1.29)
⎩ ⎭
0 0 1 0 x2 −x1
Note that Betti’s first formula (1.21) reads

3
∂
a(u, v) = − σij (u, x)vi (x)dx + γ0int v(x) γ1int u(x)dsx ,
i,j=1
∂x j
Ω Γ

3
∂
a(v, u) = − σij (v, x)ui (x)dx + γ0int u(x) γ1int v(x)dsx
i,j=1
∂xj
Ω Γ
when exchanging the role of u and v. From the symmetry of the bilinear form
a(·, ·) we then obtain Betti’s second formula

3
∂
− σij (u, x)vi (x)dx + γ0int v(x) γ1int u(x)dsx (1.30)
i,j=1
∂xj
Ω Γ

3
∂
= − σij (v, x)ui (x)dx + γ0int u(x) γ1int v(x)dsx .
i,j=1
∂x j
Ω Γ
Inserting the rigid body motions v k ∈ R into Betti’s second formula (1.30)
this gives the orthogonality

3
∂
− σij (u, x)vk,i (x)dx + γ0int v k (x) γ1int u(x)dsx = 0
i,j=1
∂xj
Ω Γ
for all v k ∈ R. Hence, for the solvability of the Neumann boundary value
problem
−µ∆u(x) − (λ + µ)grad div u(x) = f (x) for x ∈ Ω,

γ1int u(x) = g N (x) for x ∈ Γ
we have to assume the solvability conditions

v k (x) f (x)dx + γ0int v k (x) g N (x)dsx = 0 for all v k ∈ R. (1.31)
Ω Γ
Note that the solution of the Neumann boundary value problem is only unique
up to the rigid body motions (1.29). A unique solution can be defined when
considering either nodal or scaling conditions in addition.
The second order system (1.25) of linear elasticity can also be written as
a scalar partial differential equation of fourth order. By setting
λ+µ
u(x) := ∆w(x) − grad div w(x) for x ∈ Ω (1.32)
λ + 2µ
from the equilibrium equations (1.25) we obtain the scalar Bi–Laplace equa-
tion
−µ∆2 w(x) = f (x) for x ∈ Ω. (1.33)
When considering a homogeneous partial differential equation with f ≡ 0
the solution of (1.25) can be described by setting w2 ≡ w3 ≡ 0 where the
remaining component w1 = ψ is the solution of the Bi–Laplace equation
−∆2 ψ(x) = 0 for x ∈ Ω.
Then,
λ + µ ∂2
u1 (x) := ∆ψ(x) − ψ(x),
λ + 2µ ∂x21
λ+µ ∂2
u2 (x) := − ψ(x),
λ + 2µ ∂x1 ∂x2
λ+µ ∂2
u3 (x) := − ψ(x)
λ + 2µ ∂x1 ∂x3
is a solution of the homogeneous system (1.25). The function ψ is known as
Airy’s stress function.
1.2.1 Plane Elasticity
To describe problems of linear elasticity in two space dimensions one may con-
sider two different approaches. In plain stress we assume that the stress tensor
depends on two space coordinates (x1 , x2 ) only, and that the x3 –coordinates
of the stress tensor disappear:
σij (x1 , x2 , x3 ) = σij (x1 , x2 ) for i, j = 1, 2;

σ3i (x) = σi3 (x) = 0 for i = 1, 2, 3.
Applying Hooke’s law (1.19) we then obtain
e3i (u, x) = ei3 (u, x) = 0 for i = 1, 2 (1.34)
and
ν
e33 (u, x) = − [e11 (u, x) + e22 (u, x)]. (1.35)
1−ν
The resulting stress–strain relation reads
2
Eν E
σij (u, x) = δij ekk (u, x) + eij (u, x)
(1 + ν)(1 − ν) 1+ν
k=1
for x ∈ Ω and i, j = 1, 2. With
ui (x1 , x2 , x3 ) = ui (x1 , x2 ) for i = 1, 2

we further obtain from (1.34)
u3 (x1 , x2 , x3 ) = u3 (x3 ) .
In addition, (1.35) gives

∂ ν ∂ ∂
u3 (x3 ) = − u1 (x1 , x2 ) + u2 (x1 , x2 ) ,
∂x3 1 − ν ∂x1 ∂x2
and therefore

ν ∂ ∂
u3 (x) = − u1 (x) + u2 (x) x3 .
1 − ν ∂x1 ∂x2
To ensure compatibility with (1.34) we have to neglect terms of order O(x3 )

in the definition of the strain tensor eij (u, x). By introducing the modified
Lamé constants
= Eν E
λ , =
µ
(1 + ν)(1 − ν) 2(1 + ν)
we then obtain a system of partial differential equations to find the displace-

ment field (u1 , u2 ) such that
− +µ
µ∆u(x) − (λ )grad div u(x) = f (x) for x ∈ Ω ⊂ R2 .
In plain strain we assume that all components eij (u, x) of the strain tensor
depend only on the space coordinates (x1 , x2 ) and that the x3 –coordinates
vanish:
eij (u, x1 , x2 , x3 ) = eij (u, x1 , x2 ) for i, j = 1, 2;

e3i (u, x) = ei3 (u, x) = 0 for i = 1, 2, 3.
For the associated displacements we then obtain
ui (x1 , x2 , x3 ) = ui (x1 , x2 ) for i = 1, 2, u3 (x) = constant,
and the stress–strain relation reads

2
Eν E
σij (u, x) = δij ekk (u, x) + eij (u, x)
(1 + ν)(1 − 2ν) 1+ν
k=1
for i, j = 1, 2 yielding the equilibrium equations (1.25) to find the displacement

field (u1 , u2 ). Obviously,
σ3i (u, x) = σi3 (u, x) = 0 for i = 1, 2,
and
Eν
σ33 (u, x) = [e11 (u, x) + e22 (u, x)] .
(1 + ν)(1 − 2ν)
With
E
σ11 (u, x) + σ22 (u, x) = [e11 (u, x) + e22 (u, x)]
(1 + ν)(1 − 2ν)
we finally get
σ33 (u, x) = ν [σ11 (u, x) + σ22 (u, x)] .
In both cases of two–dimensional plane stress and plane strain linear elasticity
models the rigid body motions are given as

1 0 −x2
R = span , , . (1.36)
0 1 x1
For the first component of the boundary stress we obtain for x ∈ Γ

2

(γ1int u)1 (x) = σ1j (u, x)nj (x)
j=1

∂ ∂ ∂
= λ div u(x) + 2µ u1 (x) n1 (x) + µ u2 (x) + u1 (x) n2 (x)
∂x1 ∂x1 ∂x2

∂ ∂ ∂
= λ div u(x) n1 (x) + 2µ u1 (x) + µ u2 (x) − u1 (x) n2 (x),
∂nx ∂x1 ∂x2
and for the second component

∂
(γ1int u)2 (x) = λ div u(x) n2 (x) + 2µ u2 (x)
∂nx

∂ ∂
+µ u1 (x) − u2 (x) n1 (x).
∂x2 ∂x1
If we define for a two–dimensional vector field v the rotation as

∂ ∂
curl v = v2 (x) − v1 (x),
∂x1 ∂x2
and if we declare

a2
a × α := α , a ∈ R2 , α ∈ R,
−a1
we can write the boundary stress as in the representation (1.27),
∂
γ1int u(x) = [λ div u(x)]n(x) + 2µ u(x) + µ n(x) × curl v(x), x ∈ Γ.
∂nx
1.2.2 Incompressible Elasticity
For d = 2, 3 we consider the system (1.25) describing the partial differential

equations of linear elasticity. Here we are interested in almost incompressible
materials, i.e. for ν → 12 we conclude λ → ∞. Hence we introduce
p(x) := −(λ + µ) div u(x) for x ∈ Ω
to obtain from (1.25)

1
−µ∆u(x) + ∇p(x) = f (x), div u(x) = − p(x) for x ∈ Ω.
λ+µ
1
In the incompressible case ν = 2 this is equivalent to
−µ∆u(x) + ∇p(x) = f (x), div u(x) = 0 for x ∈ Ω (1.37)
which coincides with the Stokes system which plays an important role in fluid
mechanics.
1.3 Stokes System
When considering the Stokes system [91] we have to find a velocity field u and
the pressure p satisfying the system of partial differential equations
−µ ∆u(x) + ∇p(x) = f (x), div u(x) = 0 for x ∈ Ω ⊂ Rd (1.38)
where µ is the viscosity constant. If we assume Dirichlet boundary conditions

u(x) = g(x) for x ∈ Γ , integration by parts of the second equation gives

0 = div u(x) dx = [n(x)] u(x)dsx = [n(x)] g(x)dsx . (1.39)
Ω Γ Γ
Therefore, the given Dirichlet data g(x) have to satisfy the solvability condi-
tion (1.39.) Moreover, the pressure p is only unique up some additive constant.
When multiplying the components of the first partial differential equa-
tion in (1.38) with some test function vi , integrating over Ω, and applying
integration by parts this gives

∂
fi (x)vi (x)dx = −µ ∆ui (x)vi (x)dx + p(x)vi (x)dx (1.40)
∂xi
Ω Ω Ω

∂
= −µ ∆ui (x)vi (x)dx − p(x) vi (x)dx + p(x)ni (x)vi (x)dsx .
∂xi
Ω Ω Γ
1.3 Stokes System 13
Moreover we have
∂ ∂ ∂
[eij (u, x)vi (x)] = vi (x) eij (u, x) + eij (u, x) vi (x),
∂xj ∂xj ∂xj
as well as

d
∂ d
eij (u, x) vi (x) = eij (u, x)eij (v, x).
i,j=1
∂xj i,j=1
For d = 3 and i = 1 we compute
3
∂ ∂ ∂ ∂
e1j (u, x) = e11 (u, x) + e12 (u, x) + e13 (u, x)
j=1
∂xj ∂x1 ∂x2 ∂x3

∂2 1 ∂ ∂ ∂
= u 1 (x) + u 2 (x) + u 1 (x)
∂x21 2 ∂x2 ∂x1 ∂x2

1 ∂ ∂ ∂
+ u3 (x) + u1 (x)
2 ∂x3 ∂x1 ∂x3

1 1 ∂ ∂ ∂ ∂
= ∆u1 (x) + u1 (x) + u2 (x) + u3 (x)
2 2 ∂x1 ∂x1 ∂x2 ∂x3
1 1 ∂
= ∆u1 (x) + div u(x).
2 2 ∂x1
Corresponding results hold for i = 2, 3 and d = 2, respectively. Then we obtain

d
∂ d
∂ d
∂
[eij (u, x)vi (x)] = eij (u, x)vi (x) + eij (u, x) vi (x)
i,j=1
∂x j i,j=1
∂x j i,j=1
∂x j
d
1 d
∂
= ∆ui (x) + div u(x) vi (x) + eij (u, x)eij (v, x).
2 i=1 ∂xi i,j=1
This can be rewritten as

d
d
d
∂
− ∆ui (x)vi (x) = 2 eij (u, x)eij (v, x) − 2 [eij (u, x)vi (x)]
i=1 i,j=1 i,j=1
∂xj

d
∂
+ vi (x) div u(x).
i=1
∂xi
Taking the sum of (1.40) for i = 1, . . . , d, substituting the above results, and
applying integration by parts this gives

d
d
∂
v(x) f (x)dx = −µ vi (x)∆ui (x)dx + vi (x) p(x) dx
∂xi
Ω Ω i=1 Ω i=1

d
d
∂
= 2µ eij (u, x)eij (v, x)dx − 2µ [eij (u, x)vi (x)] dx
∂xj
Ω i,j=1 Ω i,j=1

d
d
∂ ∂
+µ vi (x) div u(x)dx + vi (x) p(x) dx
i=1
∂x i ∂xi
Ω Ω i=1

d
d
= 2µ eij (u, x)eij (v, x)dx − 2µ nj (x)eij (u, x)γ0int vi (x)dsx
Ω i,j=1 Γ i,j=1

−µ div u(x) div v(x) dx + µ div u(x)n(x) γ0int v(x)dsx
Ω Γ
− p(x)div v(x)dx + p(x)n(x) γ0int v(x)dsx .

Ω Γ
Hence we obtain Green’s first formula for the Stokes system

d

∂
a(u, v) = −µ∆ui (x) + p(x) vi (x)dx (1.41)
∂xi
Ω i=1

d
+ p(x)div v(x)dx + ti (u, p)vi (x)dsx
Ω Γ i=1
with the symmetric bilinear form

d
a(u, v) := 2µ eij (u, x)eij (v, x)dx − µ div u(x) div v(x) dx, (1.42)
Ω i,j=1 Ω
and with the conormal derivative

d
ti (u, p) := −[p(x) + µ div u(x)]ni (x) + 2µ eij (u, x)nj (x)
j=1
defined for x ∈ Γ and i = 1, . . . , d. For divergence–free functions u satisfying

div u = 0 we obtain, as for the system of linear elastostatics, the representation
∂
t(u, p) = −p(x) n(x) + 2µ u(x) + µ n(x) × curl u(x), x ∈ Γ. (1.43)
∂nx
Besides the standard boundary conditions (1.28) sliding boundary conditions
are often considered in fluid mechanics. In particular, for x ∈ ΓS we describe
1.4 Helmholtz Equation 15
non–penetration in normal direction and no adherence in tangential direction

by the boundary conditions
n(x) u(x) = 0, tT (x) := t(x) − [n(x) t(x)]n(x) = 0.
1.4 Helmholtz Equation
The wave equation
1 ∂2
U (x, t) = ∆U (x, t) for x ∈ Rd (d = 2, 3)
c2 ∂t2
describes the wave propagation in a homogeneous, isotropic and friction–free
medium having the constant speed of sound c. Examples are acoustic scatter-
ing and sound radiation problems.
For time harmonic acoustic waves

U (x, t) = Re u(x)e−iωt
with a frequency ω we obtain a reduced wave equation or the Helmholtz

equation
−∆u(x) − k 2 u(x) = 0 for x ∈ Rd (1.44)
where u is a scalar valued complex function and k = ω/c > 0 is the wave
number.
Let us first consider the Helmholtz equation (1.44) in a bounded domain
Ω ⊂ Rd ,
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω.
Multiplying this with a test function v, integrating over Ω, and applying
integration by parts, this gives Green’s first formula

a(u, v) = [−∆u(x) − k 2 u(x)]v(x)dx + γ1int u(x)γ0int v(x)dsx (1.45)
Ω Γ
with the symmetric bilinear form

a(u, v) = ∇u(x)∇v(x)dx − k 2 u(x)v(x)dx.
Ω Ω
Note that γ1int u = n(x) · ∇u(x) is the normal derivative of u in x ∈ Γ .

Exchanging the role of u and v we obtain in the same way

2
a(v, u) = [−∆v(x) − k v(x)]u(x)dx + γ1int v(x)γ0int u(x)dsx ,
Ω Γ
and by using the symmetry of the bilinear form a(·, ·) we conclude Green’s
second formula

2
[−∆u(x) − k u(x)]v(x)dx + γ1int u(x)γ0int v(x)dsx (1.46)
Ω Γ

= [−∆v(x) − k 2 v(x)]u(x)dx + γ1int v(x)γ0int u(x)dsx .
Ω Γ
For any solution u of the Helmholtz equation (1.44) we find from (1.45) by
setting v = u

|∇u(x)|2 dx − k 2 |u(x)|2 dx = γ1int u(x)γ0int u(x)dsx . (1.47)
Ω Ω Γ
Next we consider the Helmholtz equation (1.44) in an unbounded domain,
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω c = Rd \Ω
where we have to add the Sommerfeld radiation condition

x 1
· ∇u(x) − iku(x) = O as |x| → ∞. (1.48)
|x| |x|2
For x0 ∈ Ω let BR (x0 ) be a ball with center x0 and radius R such that
Ω ⊂ BR (x0 ) is satisfied. Then, ΩR = BR (x0 )\Ω is a bounded domain for
which we can write (1.47) as

2 2
|∇u(x)| dx − k |u(x)|2 dx
ΩR ΩR

= − γ1ext u(x)γ0ext u(x)dsx + γ1int u(x)γ0int u(x)dsx
Γ ∂BR (x0 )
taking into account the opposite direction of the normal vector on Γ . This
clearly implies

Im γ1int u(x)γ0int u(x)dsx = Im γ1ext u(x)γ0ext u(x)dsx = O(1).
∂BR (x0 ) Γ
On the other hand, from the Sommerfeld radiation condition (1.48) we also
conclude the weaker condition due to Rellich,
2
int
lim γ1 u(x) − ikγ0int u(x) dsx = 0. (1.49)
R→∞
∂BR (x0 )
1.5 Exercises 17
From this we find

2
int
0 = lim γ1 u(x) − ikγ0int u(x) dsx
R→∞
∂BR (x0 )
⎡
2 2
⎢ int 2 int
= lim ⎣ γ1 u(x) dsx + k γ0 u(x) dsx
R→∞
∂BR (x0 ) ∂BR (x0 )
⎤

⎥
−2k Im γ1int u(x)γ0int u(x)dsx ⎦
∂BR (x0 )
⎡
2 2
⎢ int 2 int
= lim ⎣ γ1 u(x) dsx + k γ0 u(x) dsx
R→∞
∂BR (x0 ) ∂BR (x0 )
⎤

−2k Im γ1ext u(x)γ0ext u(x)dsx ⎦
Γ
and therefore

2k Im γ1ext u(x)γ0ext u(x)dsx
Γ
⎡ ⎤
2 2
⎢ int int ⎥
= lim ⎣ γ1 u(x) dsx + k 2 γ0 u(x) dsx ⎦ ≥ 0
R→∞
∂BR (x0 ) ∂BR (x0 )
implying 2
int
lim γ0 u(x) dsx = O(1)
R→∞
∂BR (x0 )
as well as
1
|u(x)| = O as |x| → ∞. (1.50)
|x|
1.5 Exercises
1.1 For x ∈ R2 we consider polar coordinates
x1 = x1 (r, ϕ) = r cos ϕ, x2 = x2 (r, ϕ) = r sin ϕ for r > 0, ϕ ∈ [0, 2π).
Then, a given function u(x) can be written as
(r, ϕ).
u(x1 , x2 ) = u(x1 (r, ϕ), x2 (r, ϕ)) = u
(r, ϕ) in terms of the gradient of u(x1 , x2 ), i.e. find

Express the gradient of u
a matrix J such that
∇(r,ϕ) u
(r, ϕ) = J ∇x u(x).
Derive a representation of ∇x u(x) in terms of ∇(r,ϕ) u

(r, ϕ).
1.2 Rewrite the two–dimensional Laplace operator
∂2 ∂2
∆u(x) = u(x) + u(x)
∂x21 ∂x22
when using polar coordinates.

1.3 Prove that for any α ∈ R+ and x = 0
(r, ϕ) = rα sin(αϕ)
u(x) = u
is a solution of the two–dimensional Laplace equation.

1.4 Rewrite the three–dimensional Laplace operator
∂2 ∂2 ∂2
∆u(x) = 2 u(x) + 2 u(x) + u(x) for x ∈ R3
∂x1 ∂x2 ∂x23
when using spherical coordinates
x1 = r cos ϕ sin ϑ, x2 = r sin ϕ sin ϑ, x3 = r cos ϑ
where r > 0, ϕ ∈ [0, 2π), ϑ ∈ [0, π].

1.5 Consider the Navier system
−µ∆u(x) − (λ + µ)grad divu(x) = f (x) for x ∈ Rd .
Determine the constant α ∈ R such that the solution of the Navier system
u(x) = ∆w(x) + α grad divw(x)
can be found via the solution of a Bi–Laplace equation.

1.6 Compute all eigenvalues λk and associated eigenvectors uk of the Dirichlet
eigenvalue problem
−uk (x) = λuk (x) for x ∈ (0, 1), uk (0) = uk (1) = 0.

2
Function Spaces
In this chapter we introduce the most important function spaces as needed

for the weak formulation of boundary value problems. For a further reading
we refer to [1, 103, 106].
2.1 The Spaces C k(Ω), C k,κ(Ω) and Lp(Ω)
For d ∈ N we call a vector α = (α1 , . . . , αd ), αi ∈ N0 , multi index with the

absolute value |α| = α1 + · · · + αd and with the factorial α! = α1 ! . . . αd !. For
x ∈ Rd we can therefore write
αd
1 · · · xd .
xα = xα1
If u is a sufficient smooth real valued function, then we can write partial

derivatives as
α1 αd
α ∂ ∂
D u(x) := ... u(x1 , . . . , xd ).
∂x1 ∂xd
Let Ω ⊆ Rd be some open subset and assume k ∈ N0 . C k (Ω) is the space of

functions which are bounded and k times continuously differentiable in Ω. In
particular, for u ∈ C k (Ω) the norm

uC k (Ω) := sup |Dα u(x)|
x∈Ω
|α|≤k
is finite. Correspondingly, C ∞ (Ω) is the space of functions which are bounded

and infinitely often continuously differentiable. For a function u(x) defined for
x ∈ Ω we denote
supp u := {x ∈ Ω : u(x) = 0}
to be the support of the function u. Then,
20 2 Function Spaces
C0∞ (Ω) := {u ∈ C ∞ (Ω) : supp u ⊂ Ω}
is the space of C ∞ (Ω) functions with compact support.

For k ∈ N0 and κ ∈ (0, 1) we define C k,κ (Ω) to be the space of Hölder
continuous functions equipped with the norm
|Dα u(x) − Dα u(y)|
uC k,κ (Ω) := uC k (Ω) + sup .
x,y∈Ω,x=y |x − y|κ
|α|=k
In particular for κ = 1 we have C k,1 (Ω) to be the space of functions u ∈ C k (Ω)

where the derivatives Dα u of order |α| = k are Lipschitz continuous.
The boundary of an open set Ω ⊂ Rd is defined as
Γ := ∂Ω = Ω ∩ (Rd \Ω) .
We require that for d ≥ 2 the boundary Γ = ∂Ω can be represented locally

as the graph of a Lipschitz function using different systems of Cartesian coor-
dinates for different parts of Γ , as necessary. The simplest case occurs when
there is a function γ : Rd−1 → R such that

Ω := x ∈ Rd : xd < γ( = (x1 , . . . , xd−1 ) ∈ Rd−1 .
x) for all x
If γ(·) is Lipschitz,
|γ(
x) − γ(
y )| ≤ L |
x − y| for all x
, y ∈ Rd−1
then Ω is said to be a Lipschitz hypograph with boundary

Γ = x ∈ Rd : xn = γ( ∈ Rd−1 .
x) for all x
Definition 2.1. The open set Ω ⊂ Rd , d ≥ 2, is a Lipschitz domain if its

boundary Γ = ∂Ω is compact and if there exist finite families {Wi } and {Ωj }
having the following properties:
i. The family {Wj } is a finite open cover of Γ , i.e. Wj ⊂ Rd is an open
subset and Γ ⊆ ∪j Wj .
ii. Each Ωj can be transformed to a Lipschitz hypograph by a rigid motion,
i.e. by rotations and translations.
iii. For all j the equality Wj ∩ Ω = Wj ∩ Ωj is satisfied.
The local representation of a Lipschitz boundary Γ = ∂Ω, i.e. the choice

of families Wj and Ωj , is in general not unique. Examples for non–Lipschitz
domains are given in Fig. 2.1, see also [103].
If the parametrizations satisfy γ ∈ C k (Rd−1 ) or γ ∈ C k,κ (Rd−1 ) we call
the boundary k times differentiable or Hölder continuous, respectively. If this
holds only locally, we call the boundary piecewise smooth.
2.1 The Spaces C k (Ω), C k,κ (Ω) and Lp (Ω) 21
Fig. 2.1. Examples for non–Lipschitz domains.
By Lp (Ω) we denote the space of all equivalence classes of measurable

functions on Ω whose powers of order p are integrable. The associated norm
is ⎧ ⎫1/p
⎨ ⎬
uLp (Ω) := |u(x)|p dx for 1 ≤ p < ∞.
⎩ ⎭
Ω
Two elements u, v ∈ Lp (Ω) are identified with each other if they are different
only on a set K of zero measure µ(K) = 0. In what follows we always consider
one represent u ∈ Lp (Ω). In addition, L∞ (Ω) is the space of functions u which
are measurable and bounded almost everywhere with the norm
uL∞ (Ω) := ess sup {|u(x)|} := inf sup |u(x)| .

x∈Ω K⊂Ω,µ(K)=0 x∈Ω\K
The spaces Lp (Ω) are Banach spaces with respect to the norm · Lp (Ω) .
There holds the Minkowski inequality
u + vLp (Ω) ≤ uLp (Ω) + vLp (Ω) for all u, v ∈ Lp (Ω). (2.1)
For u ∈ Lp (Ω) and v ∈ Lq (Ω) with adjoint parameters p and q, i.e.
1 1
+ = 1,
p q
we further have Hölder’s inequality

|u(x)v(x)|dx ≤ uLp (Ω) vLq (Ω) . (2.2)
Ω
Defining the duality pairing

u, vΩ := u(x)v(x)dx,
Ω
we obtain
|u, vΩ | 1 1
vLq (Ω) = sup for 1 ≤ p < ∞, + = 1.
0=u∈Lp (Ω) uLp (Ω) p q
In particular, Lq (Ω) is the dual space of Lp (Ω) for 1 ≤ p < ∞. Moreover,

L∞ (Ω) is the dual space of L1 (Ω), but L1 (Ω) is not the dual space of L∞ (Ω).
For p = 2 we have L2 (Ω) to be the space of all square integrable functions,
and Hölder’s inequality (2.2) turns out to be the Cauchy–Schwarz inequality

|u(x)v(x)|dx ≤ uL2 (Ω) vL2 (Ω) . (2.3)
Ω
Moreover, for u, v ∈ L2 (Ω)we can define the inner product

u, vL2 (Ω) := u(x)v(x)dx
Ω
and with
u, uL2 (Ω) = u2L2 (Ω) for all u ∈ L2 (Ω)
we conclude that L2 (Ω) is a Hilbert space.
2.2 Generalized Derivatives and Sobolev Spaces
1 (Ω) we denote the space of locally integrable functions, i.e. u ∈ L1 (Ω)

By Lloc loc
is integrable with respect to any closed bounded subset K ⊂ Ω .

Example 2.2. Let Ω = (0, 1) and let u(x) = 1/x. Due to
1 1
1 1
u(x)dx = lim dx = lim ln = ∞
ε→0 x ε→0 ε
0 ε
we find u ∈ L1 (Ω). For an arbitrary closed interval K := [a, b] ⊂ (0, 1) = Ω

with 0 < a < b < 1 we obtain
b
1 b
u(x)dx = dx = ln < ∞
x a
K a
and therefore u ∈ Lloc

1 (Ω).
2.2 Generalized Derivatives and Sobolev Spaces 23
For functions ϕ, ψ ∈ C0∞ (Ω) we may apply integration by parts,

∂ ∂
ϕ(x) ψ(x) dx = − ϕ(x) ψ(x) dx.
∂xi ∂xi
Ω Ω
Note that all integrals may be defined even for non–smooth functions. This
motivates the following definition of a generalized derivative.
Definition 2.3. A function u ∈ Lloc 1 (Ω) has a generalized partial derivative

with respect to xi , if there exists a function v ∈ Lloc
1 (Ω) satisfying

∂
v(x)ϕ(x)dx = − u(x) ϕ(x)dx for all ϕ ∈ C0∞ (Ω). (2.4)
∂xi
Ω Ω
∂
Again we denote the generalized derivative by u(x) := v(x).
∂xi
The recursive application of (2.4) enables us to define a generalized partial
derivative Dα u(x) ∈ Lloc
1 (Ω) by

|α|
α
[D u(x)]ϕ(x)dx = (−1) u(x)Dα ϕ(x)dx for all ϕ ∈ C0∞ (Ω). (2.5)
Ω Ω
Example 2.4. Let u(x) = |x| for x ∈ Ω = (−1, 1). For an arbitrary ϕ ∈ C0∞ (Ω)
we have
1 0 1
∂ ∂ ∂
u(x) ϕ(x)dx = − x ϕ(x)dx + x ϕ(x)dx
∂x ∂x ∂x
−1 −1 0
0 1
0 1
= − [x ϕ(x)]−1 + ϕ(x)dx + [x ϕ(x)]0 − ϕ(x)dx
−1 0
0 1 1
= ϕ(x)dx − ϕ(x)dx = − sign(x) ϕ(x)dx
−1 0 −1
with
1 for x > 0,
sign(x) :=
−1 for x < 0.
The generalized derivative of u(x) = |x| is therefore given by
∂
u(x) = sign(x) ∈ Lloc
1 (Ω).
∂x
To compute the second derivative of u(x) = |x| we obtain
1 0 1
∂ ∂ ∂
sign(x) ϕ(x)dx = − ϕ(x)dx + ϕ(x)dx = −2ϕ(0) .
∂x ∂x ∂x
−1 −1 0
However, there exists no locally integrable function v ∈ Lloc

1 (Ω) satisfying
1
v(x)ϕ(x)dx = 2ϕ(0)
−1
for all ϕ ∈ C0∞ (Ω). Later we will find the generalized derivative of sign(x) in
the distributional sense.
For k ∈ N0 we define norms

⎧⎧ ⎫1/p
⎪
⎪⎨ ⎬
⎪
⎪
⎨ Dα upLp (Ω) for 1 ≤ p < ∞,
uWpk (Ω) := ⎩|α|≤k ⎭ (2.6)
⎪
⎪
⎪
⎪
⎩ max Dα uL∞ (Ω) for p = ∞.
|α|≤k
By taking the closure of C ∞ (Ω) with respect to the norm · Wpk (Ω) we define
the Sobolev space

·
W k (Ω)
Wpk (Ω) := C ∞ (Ω) p . (2.7)
In particular, for any u ∈ Wpk (Ω) there exists a sequence {ϕj }j∈N ⊂ C ∞ (Ω)
such that
lim u − ϕj Wpk (Ω) = 0.
j→∞
Correspondingly, the closure of C0∞ (Ω) with respect to · Wpk (Ω) defines the
Sobolev space
◦
·
W k (Ω)
W p (Ω) := C0∞ (Ω)
k p . (2.8)
The definition of Sobolev norms · Wpk (Ω) and therefore of the Sobolev spaces
(2.7) and (2.8) can be extended for any arbitrary s ∈ R. We first consider
0 < s ∈ R with s = k + κ and k ∈ N0 , κ ∈ (0, 1). Then,
! "1/p
uWps (Ω) := upW k (Ω) + |u|pW s (Ω)
p p
is the Sobolev–Slobodeckii norm, and

|Dα u(x) − Dα u(y)|p
|u|pW s (Ω) = dxdy
p |x − y|d+pκ
|α|=k Ω Ω
is the associated semi–norm. In particular for p = 2 we have W2s (Ω) to be a

Hilbert space with inner product
2.3 Properties of Sobolev Spaces 25

u, vW2k (Ω) := Dα u(x) Dα v(x) dx
|α|≤k Ω
for s = k ∈ N0 and
u, vW2s (Ω) := u, vW2k (Ω) (2.9)
(Dα u(x) − Dα u(y))(Dα v(x) − Dα v(y))

+ dxdy.
|x − y|d+2κ
|α|=k Ω Ω
for s = k + κ, k ∈ N0 , κ ∈ (0, 1).

For s < 0 and 1 < p < ∞ the Sobolev space Wps (Ω) is defined as the dual
◦
space of W q−s (Ω). Hereby we have 1/q + 1/p = 1, and the associated norm is
|u, vΩ |
uWps (Ω) := sup .
◦ vWq−s (Ω)
0=v∈W−s
q (Ω)
◦
Correspondingly, W ps (Ω) is the dual space of Wq−s (Ω).
2.3 Properties of Sobolev Spaces

In this section we state some properties of Sobolev spaces Wps (Ω) which are
needed later in the numerical analysis of finite and boundary element methods.
Assuming a certain relation for the indices s ∈ R and p ∈ N a function
u ∈ Wps (Ω) turns out to be bounded and continuous.
Theorem 2.5 (Imbedding Theorem of Sobolev). Let Ω ⊂ Rd be a

bounded domain with Lipschitz boundary ∂Ω and let
d≤s for p = 1, d/p < s for p > 1.
For u ∈ Wps (Ω) we obtain u ∈ C(Ω) satisfying
uL∞ (Ω) ≤ c uWps (Ω) for all u ∈ Wps (Ω).
For a proof of Theorem 2.5 see, for example, [31, Theorem 1.4.6], [103,
Theorem 3.26].
The norm (2.6) of the Sobolev space W21 (Ω) is
! "1/2
vW21 (Ω) = v2L2 (Ω) + ∇v2L2 (Ω)
where
|v|W21 (Ω) = ∇vL2 (Ω)
is a semi–norm. Applying the following theorem we may deduce equivalent
norms in W21 (Ω).
Theorem 2.6 (Norm Equivalence Theorem of Sobolev).

Let f : W21 (Ω) → R be a bounded linear functional satisfying
0 ≤ |f (v)| ≤ cf vW21 (Ω) for all v ∈ W21 (Ω).
If f (constant) = 0 is only satisfied for constant = 0, then

! "1/2
vW21 (Ω),f := |f (v)|2 + ∇v2L2 (Ω) (2.10)
defines an equivalent norm in W21 (Ω).
Proof. Since the linear functional f is bounded we conclude
v2W 1 (Ω),f = |f (v)|2 + ∇v2L2 (Ω)

2
≤ c2f v2W 1 (Ω) + ∇v2L2 (Ω) ≤ (1 + c2f )v2W 1 (Ω) .

2 2
The proof of the opposite direction is indirect. Assume that there is no con-
stant c0 > 0 such that
vW21 (Ω) ≤ c0 vW21 (Ω),f for all v ∈ W21 (Ω).
Then there would exist a sequence {vn }n∈N ⊂ W21 (Ω) with
vn W21 (Ω)
n ≤ for n ∈ N.
vn W21 (Ω),f
For the normalized sequence {v̄n }n∈N with

vn
v̄n :=
vn W21 (Ω)
we therefore have
vn W21 (Ω),f 1
v̄n W21 (Ω) = 1, v̄n W21 (Ω),f = ≤ →0 as n → ∞.
vn W21 (Ω) n
From this and (2.10) we conclude
lim |f (v̄n )| = 0, lim ∇v̄n L2 (Ω) = 0.

n→∞ n→∞
Since the sequence {v̄n }n∈N is bounded in W21 (Ω) and since the imbedding
W21 (Ω) → L2 (Ω) is compact [160], there exists a subsequence {v̄n }n ∈N ⊂
{v̄n }n∈N which converges in L2 (Ω). In particular, v̄ := lim

v̄n ∈ L2 (Ω).
n →∞
From
lim ∇v̄n L2 (Ω) = 0
n →∞
we obtain v̄ ∈ W21 (Ω) with ∇v̄L2 (Ω) = 0, i.e. v̄ = constant. With

2.3 Properties of Sobolev Spaces 27
0 ≤ |f (v̄)| = lim

|f (v̄n )| = 0
n →∞
we then conclude f (v̄) = 0 and therefore v̄ = 0. However, this is a contradic-

tion to
v̄W21 (Ω) = lim

v̄n W21 (Ω) = 1.

n →∞
Example 2.7. Equivalent norms in W21 (Ω) are given by

⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
vW21 (Ω),Ω := ⎣ ⎦ 2
v(x) dx + ∇vL2 (Ω) (2.11)
⎪
⎩ ⎪
⎭
Ω
and ⎧⎡ ⎫1/2
⎤2
⎪
⎨ ⎪
⎬
vW21 (Ω),Γ := ⎣ ⎦ 2
v(x) dsx + ∇vL2 (Ω) .
⎪
⎩ ⎪
⎭
Γ
◦
We therefore obtain ∇ · L2 (Ω) to be an equivalent norm in W12 (Ω)
Using the equivalent norm (2.11) as given in Example 2.7 the Poincaré
inequality
⎧⎡ ⎤2 ⎫
⎪
⎨ ⎪
⎬
2
|v(x)| dx ≤ cP ⎣ ⎦
v(x) dx + |∇v(x)| dx2
(2.12)
⎪
⎩ ⎪
⎭
Ω Ω Ω
for all v ∈ W21 (Ω) follows.

To derive some approximation properties of (piecewise) polynomial trial
spaces the following result is needed.
Theorem 2.8 (Bramble–Hilbert Lemma).

For k ∈ N0 let f : W2k+1 (Ω) → R be a bounded linear functional satisfying
|f (v)| ≤ cf vW k+1 (Ω) for all v ∈ W2k+1 (Ω).

2
By Pk (Ω) we denote the space of all polynomials of degree k defined in Ω. If
f (q) = 0
is satisfied for all q ∈ Pk (Ω) then we also have
|f (v)| ≤ c(cp ) cf |v|W k+1 (Ω) for all v ∈ W2k+1 (Ω)

2
where the constant c(cp ) depends only on the constant cp of the Poincaré
inequality (2.12).
Proof. We give only the proof for the case k = 1, i.e. P1 (Ω) is the space of
linear functions defined on Ω. For v ∈ W22 (Ω) and q ∈ P1 (Ω) we have, due to
the assumptions,
|f (v)| = |f (v) + f (q)| = |f (v + q)| ≤ cf v + qW22 (Ω) .
Moreover,
v + q2W 2 (Ω) = v + q2L2 (Ω) + |v + q|2W 1 (Ω) + |v + q|2W 2 (Ω)

2 2 2
= v + q2L2 (Ω) + ∇(v + q)2L2 (Ω) + |v|2W 2 (Ω) ,

2
since the second derivatives of a linear function disappear. Applying the

Poincaré inequality (2.12) this gives
⎡ ⎤2

v+q2W 2 (Ω) ≤ cP ⎣ [v(x) + q(x)]dx⎦ +(1+cP ) ∇(v+q)2L2 (Ω) +|v|2W 2 (Ω) .
2 2
For the second term we apply the Poincaré inequality (2.12) once again to
obtain
d 2
∂
2
∇(v + q)L2 (Ω) =
∂xi [v(x) + q(x)] dx
i=1 Ω
⎧⎡ ⎤2 ⎫
d ⎪⎨ ∂ d 2
2 ⎪⎬
⎣ ∂
≤ cP [v(x) + q(x)]dx⎦ + [v(x) + q(x)] dx
⎪ ∂xi ∂xi ∂xj ⎪
i=1 ⎩ Ω j=1 Ω
⎭
⎡ ⎤2

d
⎣ ∂
= cP [v(x) + q(x)]dx⎦ + cP |v|2W 2 (Ω)
i=1
∂xi 2
and hence
⎡ ⎤2

v + q2W 2 (Ω) ≤ cP ⎣ [v(x) + q(x)]dx⎦
2
Ω
⎡ ⎤2

d
⎣ ∂
+(1 + cP )cP [v(x) + q(x)]dx⎦ + [1 + (1 + cP )cP ] |v|2W 2 (Ω) .
i=1
∂xi 2
The assertion is proved if we can choose q ∈ P1 (Ω) so that the first two terms
are zero, i.e.

∂
[v(x) + q(x)]dx = 0, [v(x) + q(x)]dx = 0 for i = 1, . . . , d.
∂xi
Ω Ω
2.4 Distributions and Sobolev Spaces 29
With

d
q(x) = a0 + ai xi
i=1
we get
1 ∂
ai = − v(x)dx for i = 1, . . . , d
|Ω| ∂xi
Ω
# $
and therefore
1
d
a0 = − v(x) + ai xi dx.
|Ω| i=1
Ω
The proof for general k ∈ N is almost the same.

2.4 Distributions and Sobolev Spaces

As it was observed in Example 2.4 not every function in Lloc 1 (Ω) has a gener-
alized derivative in Lloc
1 (Ω). Hence we also introduce derivatives in the sense
of distributions, see also [103, 150, 156, 161].
For Ω ⊆ Rd we first define D(Ω) := C0∞ (Ω) to be the space of all test
functions.
Definition 2.9. A complex valued continuous linear form T acting on D(Ω)

is called a distribution. T is continuous on D(Ω), if ϕk → ϕ in D(Ω) always
implies T (ϕk ) → T (ϕ). The set of all distributions on D(Ω) is denoted by
D (Ω).
For u ∈ Lloc
1 (Ω) we define the distribution

Tu (ϕ) := u(x)ϕ(x)dx for ϕ ∈ D(Ω). (2.13)
Ω
Distributions of the type (2.13) are called regular. Local integrable func-

tions u ∈ Lloc
1 (Ω) can be identified with a subset of D (Ω). Hence, instead

of Tu ∈ D (Ω) we simply write u ∈ D (Ω). Nonregular distributions are called
singular. For example, the Dirac distribution for x0 ∈ Ω,
δx0 (ϕ) = ϕ(x0 ) for ϕ ∈ D(Ω),
can not be represented as in (2.13).

For the computation of the derivative of the function v(x) = sign(x) as
considered in Example 2.4 we now obtain:
Example 2.10. Using integration by parts we have for v(x) = sign(x)
1 1
∂ ∂
sign(x) ϕ(x)dx = −2ϕ(0) = − v(x) ϕ(x)dx for all ϕ ∈ D(Ω).
∂x ∂x
−1 −1
Hence the derivative of v in the distributional sense is given by

∂
v = 2 δ0 ∈ D (Ω).
∂x
As for the generalized derivative (2.5) we can define higher order deriva-
tives Dα Tu ∈ D (Ω) of a distribution Tu ∈ D(Ω) by
(Dα Tu )(ϕ) = (−1)|α| Tu (Dα ϕ) for ϕ ∈ D(Ω).
In what follows we introduce Sobolev spaces H s (Ω) which may be equivalent

to the previously introduced Sobolev spaces W2s (Ω) when some regularity
assumptions on Ω are satisfied. The definition of Sobolev spaces H s (Ω) is
based on the Fourier transform of distributions. Hence we need to introduce
first the space S(Rd ) of rapidly decreasing functions.
Definition 2.11. S(Rd ) is the space of functions ϕ ∈ C ∞ (Rd ) satisfying

ϕk, := sup |x|k + 1 |Dα ϕ(x)| < ∞ for all k, ∈ N0 .
x∈Rd |α|≤
In particular, the function ϕ and all of their derivatives decreases faster than
any polynomial.
2
Example 2.12. For the function ϕ(x) := e−|x| we have ϕ ∈ S(Rd ), but
ϕ ∈ D(Ω) = C0∞ (Rd ).
As in Definition 2.9 we can introduce the space S (Rd ) of tempered distri-

butions as the space of all complex valued linear forms T over S(Rd ).
For a function ϕ ∈ S(Rd ) we can define the Fourier transform ϕ % ∈ S(Rd ),

:= (Fϕ)(ξ) = (2π)− 2 e−ix,ξ ϕ(x)dx for ξ ∈ Rd .
d
%
ϕ(ξ) (2.14)
Rd
The mapping F : S(Rd ) → S(Rd ) is invertible and the inverse Fourier trans-
form is given by

−1 −d
%
(F ϕ)(x) = (2π) 2 %
eix,ξ ϕ(ξ)dξ for x ∈ Rd . (2.15)
Rd
% ∈ D(Rd ).
In general, ϕ ∈ D(Rd ) does not imply ϕ
For ϕ ∈ S(Rd ) we have
Dα (Fϕ)(ξ) = (−i)|α| F(xα ϕ)(ξ) (2.16)
as well as
ξ α (Fϕ)(ξ) = (−i)|α| F(Dα ϕ)(ξ). (2.17)
Lemma 2.13. The Fourier transform maintains rotational symmetries, i.e.

for u ∈ S(Rd ) we have u
%(ξ) = u
%(|ξ|) for all ξ ∈ Rd iff u(x) = u(|x|) for all
x∈R . d
Proof. Let us first consider the two–dimensional case d = 2. Using polar

coordinates, & ' & '
|ξ| cos ψ r cos φ
ξ= , x= ,
|ξ| sin ψ r sin φ
we obtain
∞ 2π
1
%(ξ) = u
u %(|ξ|, ψ) = e−ir |ξ|[cos φ cos ψ+sin φ sin ψ] u(r)rdφdr
2π
0 0
∞ 2π
1
= e−ir|ξ| cos(φ−ψ) u(r)rdφdr.
2π
0 0
With ψ0 ∈ [0, 2π) and substituting φ := φ − ψ0 it follows that
∞ 2π
1
%(|ξ|, ψ + ψ0 ) =
u e−ir|ξ| cos(φ−ψ−ψ0 ) u(r)rdφdr
2π
0 0
∞ 2π−ψ
0
1
= e−ir|ξ| cos(φ−ψ) u(r)rdφdr.
2π
0 −ψ0
By using
0 2π

−ir|ξ| cos(φ−ψ)
e dφ = e−ir|ξ| cos(φ−ψ) dφ
−ψ0 2π−ψ0
we then obtain
%(|ξ|, ψ) = u
u %(|ξ|, ψ + ψ0 ) for all ψ0 ∈ [0, 2π)
%(ξ) = u
and therefore the assertion u %(|ξ|).
For d = 3 we use spherical coordinates

⎛ ⎞ ⎛ ⎞
|ξ| cos ψ sin ϑ r cos φ sin θ
⎜ ⎟ ⎜ ⎟
ξ = ⎝ |ξ| sin ψ sin ϑ ⎠ , x = ⎝ r sin φ sin θ ⎠
|ξ| cos ϑ r cos θ
to obtain
%(|ξ|, ψ, ϑ) =
u
∞ 2ππ
1
= e−ir|ξ|[cos(φ−ψ) sin θ sin ϑ+cos θ cos ϑ] u(r)r2 sin θdθdφdr.
(2π)3/2
0 0 0
As for the two–dimensional case d = 2 we conclude
%(|ξ|, ψ + ψ0 , ϑ) = u
u %(|ξ|, ψ, ϑ) for all ψ0 ∈ [0, 2π).
For a fixed ϑ ∈ [0, π] and for a given radius we also have u

%(ξ) = u
%(|ξ|) = u
%()
along the circular lines
ξ12 + ξ22 = 2 sin2 ϑ, ξ3 = cos ϑ.
Using permutated spherical coordinates we also find u %(ξ) = u

%() along the
circular lines
ξ12 + ξ32 = 2 sin2 ϑ, ξ2 = cos ϑ .

For a distribution T ∈ S (Rd ) we can define the Fourier transform T% ∈ S (Rd )
T%(ϕ) := T (ϕ)
% for ϕ ∈ S(Rd ).
The mapping F : S (Rd ) → S (Rd ) is invertible and the inverse Fourier trans-
form is given by
(F −1 T )(ϕ) := T (F −1 ϕ) for ϕ ∈ S(Rd ).
The rules (2.16) and (2.17) remain valid for distributions T ∈ S (Rd ).
For s ∈ R and u ∈ S(Rd ) we define the Bessel potential operator

−d/2
J u(x) := (2π)
s
(1 + |ξ|2 )s/2 u
%(ξ)eix,ξ dξ, x ∈ Rd ,
Rd
which is a bounded linear operator J s : S(Rd ) → S(Rd ). The application of

the Fourier transform gives
(FJ s u)(ξ) = (1 + |ξ|2 )s/2 (Fu)(ξ) .

From this we conclude that the application of J s corresponds in the Fourier

space to a multiplication with a function of order O(|ξ|s ). Therefore, using
(2.17) we can see J s as a differential operator of order s.
For T ∈ S (Rd ) we define a bounded and linear operator J s : S (Rd ) →
S (Rd ) acting on the space of tempered distributions,
(J s T )(ϕ) := T (J s ϕ) for all ϕ ∈ S(Rd ).
The Sobolev space H s (Rd ) is the space of all distributions v ∈ S (Rd ) with
J s v ∈ L2 (Rd ) where the associated inner product
u, vH s (Rd ) := J s u, J s vL2 (Rd )
implies the norm

u2H s (Rd ) := J s u||2L2 (Rd ) = (1 + |ξ|2 )s |%
u(ξ)|2 dξ .
Rd
The connection with the Sobolev spaces W2s (Rd ) can be seen from the follow-
ing theorem, see for example [103, 160].
Theorem 2.14. For all s ∈ R there holds
H s (Rd ) = W2s (Rd ) .
For a bounded domain Ω ⊂ Rd we define the Sobolev space H s (Ω) by

restriction,
H s (Ω) := v = v|Ω : v ∈ H s (Rd ) ,
with the norm
vH s (Ω) := inf
v H s (Rd ) .
∈H s (Rd ),
v v|Ω =v
In addition we introduce Sobolev spaces
s (Ω) := C ∞ (Ω)
·
H s (Rd ) ,
H H0s (Ω) := C0∞ (Ω)

·
H s (Ω)
0
which will coincide for almost all s ∈ R+ , see for example [103, Theorem 3.33].
Theorem 2.15. Let Ω ⊂ Rd be a Lipschitz domain. For s ≥ 0 we have
s (Ω) ⊂ H s (Ω).
H 0
In particular,

s (Ω) = H s (Ω) 1 3 5
H 0 for s ∈ , , ,... .
2 2 2
Moreover,
s (Ω) = [H −s (Ω)] ,
H −s (Ω)]
H s (Ω) = [H for all s ∈ R.
The equivalence of Sobolev spaces W2s (Ω) and H s (Ω) holds only when
certain assumptions on Ω are satisfied. Sufficient for the norm equivalence is
the existence of a linear bounded extension operator
EΩ : W2s (Ω) → W2s (Rd ).
This condition is ensured for a bounded domain Ω ⊂ Rd , if a uniform cone
condition is satisfied, see for example [161, Theorem 5.4].
Theorem 2.16. For a Lipschitz domain Ω ⊂ Rd we have
W2s (Ω) = H s (Ω) for all s > 0.
For the analysis of the Stokes system we need to have some mapping
properties of the gradient ∇.
Theorem 2.17. [53, Theorem 3.2, p. 111] Let the Lipschitz domain Ω ⊂ Rd
be bounded and connected. Then there holds

qL2 (Ω) ≤ c1 qH −1 (Ω) + ∇q[H −1 (Ω)]d for all q ∈ L2 (Ω)
as well as

qL2 (Ω) ≤ c2 ∇q[H −1 (Ω)]d for all q ∈ L2 (Ω) with q(x)dx = 0. (2.18)
Ω
Bounded linear operators can be seen as maps between different Sobolev

spaces inducing different operator norms. Then one can extend the bound-
edness properties to Sobolev spaces between. For a general overview on in-
terpolation spaces we refer to [16, 103] . Here we will use only the following
result.
Theorem 2.18 (Interpolation Theorem). Let A : H α1 (Ω) → H β (Ω) be
some bounded and linear operator with norm
AvH β (Ω)
Aα1 ,β := sup .
0=v∈H α1 (Ω) vH α1 (Ω)
For α2 > α1 let A : H α2 (Ω) → H β (Ω) be bounded with norm Aα2 ,β . Then

the operator A : H α (Ω) → H β (Ω) is bounded for all α ∈ [α1 , α2 ] and the
corresponding operator norm is given by
α−α2 α−α1
Aα,β ≤ (Aα1 ,β ) α1 −α2 (Aα2 ,β ) α2 −α1 .
Let the operator A : H α (Ω) → H β1 (Ω) be bounded with norm Aα,β1 and
let A : H α (Ω) → H β2 (Ω) be bounded with norm Aα,β2 assuming β1 < β2 .
Then the operator A : H α (Ω) → H β (Ω) is bounded for all β ∈ [β1 , β2 ] and
the corresponding operator norm is given by
β−β2 β−β1
Aα,β ≤ (Aα,β1 ) β1 −β2 (Aα,β2 ) β2 −β1 .
2.5 Sobolev Spaces on Manifolds 35
2.5 Sobolev Spaces on Manifolds
Let Ω ⊂ Rd be a bounded domain (d = 2, 3) and let the boundary Γ = ∂Ω

be given by some arbitrary overlapping piecewise parametrization
*
J

Γ = Γi , Γi := x ∈ Rd : x = χi (ξ) for ξ ∈ Ti ⊂ Rd−1 . (2.19)
i=1
With respect to (2.19) we also consider a partition of unity, {ϕi }pi=1 , of non–
negative cut off functions ϕi ∈ C0∞ (Rd ) satisfying

J
ϕi (x) = 1 for x ∈ Γ, ϕi (x) = 0 for x ∈ Γ \Γi .
i=1
For any function v defined on the boundary Γ we can write

J
J
v(x) = ϕi (x)v(x) = vi (x) for x ∈ Γ
i=1 i=1
with vi (x) := ϕi (x)v(x). Inserting the local parametrizations (2.19) we obtain

for i = 1, . . . , J
vi (x) = ϕi (x)v(x) = ϕi (χi (ξ))v(χi (ξ)) =: vi (ξ) for ξ ∈ Ti ⊂ Rd−1 .
The functions vi are defined with respect to the parameter domains Ti ⊂ Rd−1
for which we can introduce appropriate Sobolev spaces. Taking into account
the chain rule we have to ensure the existence of all corresponding derivatives
of the local parametrization χi (ξ). For the definition of derivatives of order
|s| ≤ k we therefore have to assume χi ∈ C k−1,1 (Ti ). In particular for a
Lipschitz domain with a local parametrization χi ∈ C 0,1 (Ti ) we can only
introduce Sobolev spaces H s (Ti ) for |s| ≤ 1.
In general we can define the Sobolev norm
J +1/2

vHχs (Γ ) := vi 2H s (Ti )
(2.20)
i=1
for 0 ≤ s ≤ k and therefore the corresponding Sobolev space H s (Γ ).
Lemma 2.19. For s = 0 an equivalent norm in Hχ0 (Γ ) is given by

⎧ ⎫1/2
⎨ ⎬
vL2 (Γ ) := |v(x)|2 dsx .
⎩ ⎭
Γ
Proof. First we note that

J

v2Hχ0 (Γ ) = [ϕi (χi (ξ))v(χi (ξ))]2 dξ
i=1 T
i
and
J

v2L2 (Γ ) = [v(x)]2 dsx = ϕi (x)[v(x)]2 dsx .
Γ i=1 Γ
i
Inserting the local parametrization this gives

J

v2L2 (Γ ) = ϕi (χi (ξ))[v(χi (ξ))]2 detχi (ξ)dξ.
i=1 T
i
From this the assertion follows, where the constants depend on both the chosen
parametrization (2.19) and on the particular definition of the cut off functions
ϕi .

For s ∈ (0, 1) we find in the same way that the Sobolev–Slobodeckii norm
⎧ ⎫1/2
⎨ 2 ⎬
[v(x) − v(y)]
vH s (Γ ) := v2L2 (Γ ) + ds ds
x y
⎩ |x − y|d−1+2s ⎭
Γ Γ
is an equivalent norm in Hχs (Γ ).

As in the Equivalence Theorem of Sobolev (Theorem 2.6) we may also
define other equivalent norms in H s (Γ ). For example,
⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
[v(x) − v(y)]2
vH 1/2 (Γ ),Γ := ⎣ v(x)dsx ⎦ + ds ds
x y
⎪
⎩ |x − y|d ⎪
⎭
Γ Γ Γ
defines an equivalent norm in H 1/2 (Γ ).

Up to now we only considered Sobolev spaces H s (Γ ) for s ≥ 0. For s < 0
H (Γ ) is defined as the dual space of H −s (Γ ),
s
H s (Γ ) := [H −s (Γ )] ,
where the associated norm is

w, vΓ
wH s (Γ ) := sup
0=v∈H −s (Γ ) vH −s (Γ )
with respect to the duality pairing

w, vΓ := w(x)v(x)dsx .
Γ
2.5 Sobolev Spaces on Manifolds 37
Let Γ0 ⊂ Γ be some open part of a sufficient smooth boundary Γ = ∂Ω. For

s ≥ 0 we introduce the Sobolev space

H s (Γ0 ) := v = v|Γ0 : v ∈ H s (Γ )
with the norm

vH s (Γ0 ) := inf
v H s (Γ )
∈H s (Γ ):
v v|Γ0 =v
as well as the Sobolev space

H s (Γ0 ) := v = v|Γ : v ∈ H s (Γ ), supp v ⊂ Γ0 .
0
For s < 0 we define the appropriate Sobolev spaces by duality,

−s (Γ0 )] ,
H s (Γ0 ) := [H s (Γ0 ) := [H −s (Γ0 )] .
H
Finally we consider a closed boundary Γ = ∂Ω which is piecewise smooth,
*
J
Γ = Γ i, Γi ∩ Γj = ∅ for i = j.
i=1
For s > 0 we define by

s
Hpw (Γ ) := v ∈ L2 (Γ ) : v|Γi ∈ H s (Γi ), i = 1, . . . , J
the space of piecewise smooth functions with the norm

J +1/2

vHpw
s (Γ ) := v|Γi 2H s (Γi )
i=1
while for s < 0 we have
,
J
s
Hpw (Γ ) := s (Γj )
H (2.21)
j=1
with the norm

J
wHpw
s (Γ ) := w|Γj H
-s (Γj ) . (2.22)
j=1
Lemma 2.20. For w ∈ Hpw

s
(Γ ) and s < 0 we have
wH s (Γ ) ≤ wHpw
s (Γ ) .
Proof. By duality we conclude
|w, vΓ | J
|w, vΓj |
wH s (Γ ) = sup ≤ sup
−s
0=v∈H (Γ ) v H −s (Γ ) −s
0=v∈H (Γ ) j=1 vH −s (Γ )

J
|w|Γj , v|Γj Γj |
≤ sup
0=v∈H −s (Γ ) j=1 v|Γj H −s (Γj )

J
|w|Γj , vj Γj |
≤ sup = wHpw
s (Γ ) .

j=1 0=vj ∈H −s (Γ j)
vj H −s (Γj )
If Γ = ∂Ω is the boundary of a Lipschitz domain Ω ⊂ Rd , then we have

to assume |s| ≤ 1 to ensure the above statements.
For a function u given in a bounded domain Ω ⊂ Rd the application of the
interior trace (1.3) gives γ0int u as a function on the boundary Γ = ∂Ω. The
relations between the corresponding function spaces are stated in the next
two theorems, see, for example, [1, 103, 160].
Theorem 2.21 (Trace Theorem). Let Ω ⊂ Rd be a C k−1,1 –domain. For
1
2 < s ≤ k the interior trace operator
γ0int : H s (Ω) → H s−1/2 (Γ )

is bounded satisfying
γ0int vH s−1/2 (Γ ) ≤ cT vH s (Ω) for all v ∈ H s (Ω).
For a Lipschitz domain Ω we can apply Theorem 2.21 with k = 1 to
obtain the boundedness of the trace operator γ0int : H s (Ω) → H s−1/2 (Γ ) for
s ∈ ( 12 , 1]. This remains true for s ∈ ( 12 , 32 ), see [44] and [103, Theorem 3.38].
Theorem 2.22 (Inverse Trace Theorem). Let Ω be a C k−1,1 –domain.
For 12 < s ≤ k the interior trace operator γ0int : H s (Ω) → H s−1/2 (Γ ) has a
continuous right inverse operator
E : H s−1/2 (Γ ) → H s (Ω)
satisfying γ0int Ew = w for all w ∈ H s−1/2 (Γ ) as well as

EwH s (Ω) ≤ cIT wH s−1/2 (Γ ) for all w ∈ H s−1/2 (Γ ).
Therefore, for s > 0 we can define Sobolev spaces H s (Γ ) also as trace
spaces of H s+1/2 (Ω). The corresponding norm is given by
vH s (Γ ),γ0 := inf ||V ||H 1/2+s (Ω) .
V ∈H s+1/2 (Ω),γ0int V =v
However, for a Lipschitz domain Ω ⊂ Rd the norms vH s (Γ ),γ0 and vH s (Γ )

are only equivalent for |s| ≤ 1.
2.6 Exercises 39
Remark 2.23. The interpolation theorem (Theorem 2.18) holds also for appro-
priate Sobolev spaces H s (Γ ).
As in Theorem 2.8 we also have the Bramble–Hilbert lemma:
Theorem 2.24. Let Γ = ∂Ω the boundary of a C k−1,1 –domain Ω ⊂ Rd and

let f : H k+1 (Γ ) → R be a bounded linear functional satisfying
|f (v)| ≤ cf vH k+1 (Γ ) for all v ∈ H k+1 (Γ ).
If
f (q) = 0
is satisfied for all q ∈ Pk (Γ ) then we also have
|f (v)| ≤ c cf |v|H k+1 (Γ ) for all v ∈ H k+1 (Γ ).
2.6 Exercises
2.1 Let u(x), x ∈ (0, 1), be a continuously differentiable function satisfying

u(0) = u(1) = 0. Prove
1 1
2
[u(x)] dx ≤ c [u (x)]2 dx
0 0
where c should be as small as possible.

2.2 Consider the function

0 for x ∈ [0, 12 ],
u(x) =
1 for x ∈ ( 12 , 1].
Determine those values of s ∈ (0, 1) such that
1 1
[u(x) − u(y)]2
dx dy < ∞
|x − y|1+2s
0 0
is finite.
3
Variational Methods
The weak formulation of boundary value problems leads to variational prob-

lems and associated operator equations. In particular, the representation of
solutions of partial differential equations by using surface and volume po-
tentials requires the solution of boundary integral operator equations to find
the complete Cauchy data. In this chapter we describe the basic tools from
functional analysis which are needed to investigate the unique solvability of
operator equations.
3.1 Operator Equations
. space with the inner product ·, ·X and with the induced
Let X be a Hilbert
norm · X = ·, ·X . Let X be the dual space of X with respect to the
duality pairing ·, ·. Then it holds that
|f, v|
f X = sup for all f ∈ X . (3.1)
0=v∈X vX
Let A : X → X be a bounded linear operator satisfying
AvX ≤ cA
2 vX for all v ∈ X. (3.2)
We assume that A is self–adjoint, i.e., we have
Au, v = u, Av for all u, v ∈ X.
For a given f ∈ X we want to find the solution u ∈ X of the operator equation
Au = f . (3.3)
Instead of the operator equation (3.3) we may consider an equivalent varia-

tional problem to find u ∈ X such that
42 3 Variational Methods
Au, v = f, v for all v ∈ X. (3.4)
Obviously, any solution u ∈ X of the operator equation (3.3) is also a solution

of the variational problem (3.4). To show the reverse direction we now consider
u ∈ X to be a solution of the variational problem (3.4). Using the norm
definition (3.1) we then obtain
|Au − f, v|
Au − f X = sup = 0
0=v∈X vX
and therefore 0 = Au − f ∈ X , i.e., u ∈ X is a solution of the operator

equation (3.3).
The operator A : X → X induces a bilinear form
a(u, v) := Au, v for all u, v ∈ X
with the mapping property
a(·, ·) : X × X → R. (3.5)
In the reverse case, any bilinear form (3.5) defines an operator A : X → X .
Lemma 3.1. Let a(·, ·) : X × X → R be a bounded bilinear form satisfying
2 uX vX
|a(u, v)| ≤ cA for all u, v ∈ X.
For any u ∈ X there exists an element Au ∈ X such that
Au, v = a(u, v) for all v ∈ X.
The operator A : X → X is linear and bounded satisfying
AuX ≤ cA
2 uX for all u ∈ X.
Proof. For a given u ∈ X we define fu , v := a(u, v) which is a bounded

linear form in X, i.e., we have fu ∈ X . The map u ∈ X → fu ∈ X defines a
linear operator A : X → X with Au = fu ∈ X and satisfying
|fu , v| |a(u, v)|

AuX = fu X = sup = sup ≤ cA
2 uX .

0=v∈X vX 0=v∈X vX
If A : X → X is a self–adjoint and positive semi–definite operator we can

derive a minimization problem which is equivalent to the variational formu-
lation (3.4).
Lemma 3.2. Let A : X → X be self–adjoint and positive semi–definite, i.e.,
Av, v ≥ 0 for all v ∈ X.

3.1 Operator Equations 43
Let F be the functional

1
F (v) := Av, v − f, v for v ∈ X .
2
The solution of the variational formulation (3.4) is then equivalent to the
solution of the minimization problem
F (u) = min F (v) . (3.6)

v∈X
Proof. For u, v ∈ X we choose an arbitrary t ∈ R. Then we have

1
F (u + tv) = A(u + tv), u + tv − f, u + tv
2
1
= F (u) + t [Au, v − f, v] + t2 Av, v.
2
If u ∈ X is a solution of the variational problem (3.4) we then obtain
1
F (u) ≤ F (u) + t2 Av, v = F (u + tv)
2
for all v ∈ X and t ∈ R. Therefore, u ∈ X is also a solution of the minimization
problem (3.6).
Let u ∈ X be now a solution of (3.6). Then, as a necessary condition,
d
F (u + tv)|t=0 = 0 for all v ∈ X.
dt
From this we obtain
Au, v = f, v for all v ∈ X
and therefore the equivalence of both the variational and the minimization
problem.
To investigate the unique solvability of the operator equation (3.3) we now
consider a fixed point iteration. For this we need to formulate the following
Riesz representation theorem.
Theorem 3.3 (Riesz Representation Theorem). Any linear and bounded

functional f ∈ X can be written as
f, v = u, vX
where u ∈ X is uniquely determined by f ∈ X , and
uX = f X . (3.7)
Proof. Let f ∈ X be arbitrary but fixed. Then we can find u ∈ X as the

solution of the variational problem
u, vX = f, v for all v ∈ X. (3.8)
Using Lemma 3.2 this variational problem is equivalent to the minimization

problem
F (u) = min F (v) (3.9)
v∈X
where the functional is given by

1
F (v) = v, vX − f, v for v ∈ X.
2
Hence we have to investigate the unique solvability of the minimization prob-
lem (3.9). From
1 1
F (v) = v, vX − f, v ≥ v2X − f X vX
2 2
1 2 1 1
= [vX − f X ] − f 2X ≥ − f 2X
2 2 2
we find that F (v) is bounded below for all v ∈ X. Hence there exists the
infimum
α := inf F (v) ∈ R.
v∈X
Let {uk }k∈N ⊂ X be a sequence approaching the minimum, i.e., F (uk ) → α

as k → ∞. With the identity

uk − u 2X + uk + u 2X = 2 uk 2X + u 2X
we then obtain
0 ≤ uk − u 2X = 2 uk 2X + 2 u 2X − uk + u 2X

1 2 1 2
=4 uk X − f, uk + 4 u X − f, u
2 2
+4 f, uk + u − uk + u 2X

1
= 4 F (uk ) + 4 F (u ) − 8 F (uk + u )
2
≤ 4 F (uk ) + 4 F (u ) − 8α → 0 as k, → ∞.
Therefore, {uk }k∈N is a Cauchy sequence, and since X is a Hilbert space, we

find the limit
u = lim uk ∈ X.
k→∞
3.1 Operator Equations 45
Moreover,
1
|F (uk ) − F (u)| ≤ |uk , uk X − u, uX | + |f, uk − u|
2
1
= |uk , uk − uX + u, uk − uX | + |f, uk − u|
2

1 1
≤ uk X + uX + f X uk − uX ,
2 2
and hence
F (u) = lim F (uk ) = α.
k→∞
In particular, u ∈ X is a solution of the minimization problem (3.9) and

therefore also a solution of the variational problem (3.8).
It remains to prove the uniqueness. Let u ∈ X be another solution of (3.9)
and (3.8), respectively. Then,

u, vX = f, v for all v ∈ X.
Subtracting this from (3.8) this gives
u − u
, vX = 0 for all v ∈ X.
Choosing v = u − u
we now obtain
2X = 0
u − u
, i.e., u ∈ X is the unique solution of (3.8) and (3.9),

and therefore u = u
respectively.
Finally,
u2X = u, uX = f, u ≤ f X uX
and
|f, v| |u, vX |
f X = sup = sup ≤ uX ,
imply the norm equality (3.7).

The map J : X → X as introduced in Theorem 3.3 is called the Riesz
map u = Jf and satisfies the variational problem
Jf, vX = f, v for all v ∈ X. (3.10)
Moreover,
Jf X = f X . (3.11)
3.2 Elliptic Operators
To ensure the unique solvability of the operator equation (3.3) and of the vari-
ational problem (3.4) we need to have a further assumption for the operator
A and for the bilinear form a(·, ·), respectively. The operator A : X → X is
called X–elliptic if
2
Av, v ≥ cA
1 vX for all v ∈ X (3.12)
is satisfied with some positive constant cA

1.
Theorem 3.4 (Lax–Milgram Lemma). Let the operator A : X → X be

bounded and X–elliptic. For any f ∈ X there exists a unique solution of the
operator equation (3.3) satisfying the estimate
1
uX ≤ f X .
cA
1
Proof. Let J : X → X be the Riesz operator as defined by (3.10). The

operator equation (3.3) is then equivalent to the fixed point equation
u = u − J(Au − f ) = T u + Jf
with the operator

T := I − JA : X → X
and with a suitable chosen parameter 0 < ∈ R. From the boundedness
estimate (3.2) and from the ellipticity assumption (3.12) of A as well as from
the properties (3.10) and (3.11) of the Riesz map J we conclude
2
JAv, vX = Av, v ≥ cA
1 vX , JAvX = AvX ≤ cA
2 vX
and therefore
T v2X = (I − JA)v2X
= v2X − 2JAv, vX + 2 JAv2X

2 A 2 2
≤ [1 − 2cA
1 + (c2 ) ] vX .
A 2
Hence we obtain that for ∈ (0, 2cA
1 /(c2 ) ) the operator T is a contraction
in X, and the unique solvability of (3.3) follows from Banach’s fixed point
theorem [163]. Let u ∈ X be the unique solution of the operator equation
(3.3). Then,
2
1 uX ≤ Au, u = f, u ≤ f X uX ,
cA
which is equivalent to the remaining bound.

3.2 Elliptic Operators 47
Applying Theorem 3.4 this gives the inverse operator A−1 : X → X and
we obtain
1
A−1 f X ≤ A f X for all f ∈ X . (3.13)
c1
From the boundedness of the self–adjoint and invertible operator A we also
conclude an ellipticity estimate for the inverse operator A−1 .
Lemma 3.5. Let A : X → X be bounded, self–adjoint and X–elliptic. In

particular we assume (3.2), i.e.,
AvX ≤ cA
2 vX for all v ∈ X.
Then,
1
A−1 f, f ≥ f 2X for all f ∈ X .
cA
2
Proof. Let us consider the operator B := JA : X → X satisfying
BvX = JAvX = AvX ≤ cA
Since
Bu, vX = JAu, vX = Au, v = u, Av = u, BvX
holds for all u, v ∈ X the operator B is self–adjoint satisfying the ellipticity
estimate
2
Bv, vX = Av, v ≥ cA1 vX for all v ∈ X.
Hence there exists a self–adjoint and invertible operator B 1/2 satisfying
B = B 1/2 B 1/2 , see, e.g., [118]. In addition we define B −1/2 := (B 1/2 )−1 .
Then we obtain
B 1/2 v2X = Bv, vX ≤ BvX vX ≤ cA 2

2 vX for all v ∈ X
and further /
B 1/2 vX ≤ 2 vX
cA for all v ∈ X.
For an arbitrary f ∈ X we then conclude
|f, v| |Jf, vX | |B −1/2 Jf, B 1/2 vX |

f X = sup = sup = sup
0=v∈X vX 0=v∈X vX 0=v∈X vX
−1/2 1/2 /
B Jf X B vX −1/2
≤ sup ≤ cA 2 B Jf X ,
0=v∈X v X
and therefore
f 2X ≤ cA
2 B
−1/2
Jf 2X = cA
2 B
−1
Jf, Jf X = cA
2 A
−1
f, f .

3.3 Operators and Stability Conditions

Let Π be a Banach space and let B : X → Π be a bounded linear operator
satisfying
BvΠ ≤ cB2 vX for all v ∈ X. (3.14)
The operator B implies a bounded bilinear form b(·, ·) : X × Π → R,
b(v, q) := Bv, q for (v, q) ∈ X × Π.
The null space of the operator B is
ker B := {v ∈ X : Bv = 0} .
The orthogonal complement of ker B in X is given as
(ker B)⊥ := {w ∈ X : w, vX = 0 for all v ∈ ker B} ⊂ X .
Finally,
(ker B)0 := {f ∈ X : f, v = 0 for all v ∈ ker B} ⊂ X (3.15)
is the polar space which is induced by ker B.
For a given g ∈ Π we want to find solutions u ∈ X of the operator
equation
Bu = g. (3.16)
Obviously we have to require the solvability condition
g ∈ ImX B := {Bv ∈ Π for all v ∈ X} . (3.17)

Let B : Π → X the adjoint of B : X → Π , i.e.
v, B q := Bv, q for all (v, q) ∈ X × Π.
Then we have
ker B := {q ∈ Π : Bv, q = 0 for all v ∈ X} ,
(ker B )⊥ := {p ∈ Π : p, qΠ = 0 for all q ∈ ker B } ,
(ker B )0 := {g ∈ Π : g, q = 0 for all q ∈ ker B } .
To characterize the image ImX B we will use the following result, see, for
example, [163].
Theorem 3.6 (Closed range theorem). Let X and Π be Banach spaces,
and let B : X → Π be a bounded linear operator. Then the following proper-
ties are all equivalent:
i. ImX B is closed in Π .
ii. ImΠ B is closed in X .
iii. ImX B = (ker B )0 .
iv. ImΠ B = (ker B)0 .
3.3 Operators and Stability Conditions 49
Proof. Here we only prove that iii. follows from i., see also [163]. From the
definition of the polar space with respect to B we find that
(ker B )0 = {g ∈ Π : g, q = 0 for all q ∈ ker B }

= {g ∈ Π : g, q = 0 for all q ∈ Π : Bv, q = 0 for all v ∈ X}
and therefore
ImX B ⊂ (ker B )0 .
Let g ∈ (ker B )0 with g ∈ ImX B. Applying the separation theorem for closed
convex sets there exists a q̄ ∈ Π and a constant α ∈ R such that
g, q̄ > α > f, q̄ for all f ∈ ImX (B) ⊂ Π .
Since B is linear we obtain for an arbitrary given f ∈ ImX B also −f ∈ ImX B

and therefore
α > −f, q̄.
From this we obtain α > 0 as well as |f, q̄| < α. For f ∈ ImΠ B and for any
arbitrary n ∈ N we also conclude nf ∈ ImΠ B and therefore
α
|f, q̄| < for all n ∈ N
n
which is equivalent to
f, q̄ = 0 for all f ∈ ImX (B).
For any f ∈ ImX (B) there exists at least one u ∈ X with f = Bu. Hence,
0 = f, q̄ = Bu, q̄ = u, B q̄ for all u ∈ X,
and therefore q̄ ∈ ker B . On the other hand, for g ∈ (ker B)0 we have
g, q = 0 for all q ∈ ker B
and therefore g, q̄ = 0 which is a contradiction to g, q̄ > α > 0.

The solvability condition (3.17) is equivalent to
g, q = 0 for all q ∈ ker B ⊂ Π . (3.18)
If the equivalent solvability conditions (3.17) and (3.18) are satisfied, then
there exists at least one solution u ∈ X satisfying Bu = g. When the null
space ker B is non–trivial, we can add an arbitrary u0 ∈ ker B, in particular,
u + u0 is still a solution of B(u + u0 ) = g. In this case, the solution is not
unique in general. Instead we consider only solutions u ∈ (ker B)⊥ . To ensure
unique solvability in this case, we have to formulate additional assumptions.
Theorem 3.7. Let X and Π be Hilbert spaces and let B : X → Π be a

bounded linear operator. Further we assume the stability condition
Bv, q
cS vX ≤ sup for all v ∈ (ker B)⊥ . (3.19)
0=q∈Π qΠ
For a given g ∈ ImX (B) there exists a unique solution u ∈ (ker B)⊥ of the
operator equation Bu = g satisfying
1
uX ≤ gΠ .
cS
Proof. Since we assume g ∈ ImX B there exists at least one solution
u ∈ (ker B)⊥ of the operator equation Bu = g satisfying
Bu, q = g, q for all q ∈ Π.
Let ū ∈ (ker B)⊥ be a second solution satisfying
B ū, q = g, q for all q ∈ Π.
Then,
B(u − ū), q = 0 for all q ∈ Π.
Obviously, u − ū ∈ (ker B)⊥ . From the stability condition (3.19) we then
conclude
B(u − ū), q
0 ≤ cS u − ūX ≤ sup = 0
0=q∈Π qΠ
and therefore uniqueness, u = ū. Applying (3.19) for the solution u this gives
Bu, q g, q
cS uX ≤ sup = sup ≤ gΠ .

0=q∈Π qΠ 0=q∈Π qΠ
3.4 Operator Equations with Constraints

In many applications we have to solve an operator equation Au = f where
the solution u has to satisfy an additional constraint Bu = g. In this case we
have to assume first the solvability condition (3.17). For a given g ∈ Π we
then define the manifold
Vg := {v ∈ X : Bv = g} .
In particular, V0 = ker B. Further, the given f ∈ X has to satisfy the solv-

ability condition
f ∈ ImVg A := {Av ∈ X for all v ∈ Vg } .

3.4 Operator Equations with Constraints 51
Then we have to find u ∈ Vg satisfying the variational problem
Au, v = f, v for all v ∈ V0 . (3.20)
The unique solvability of (3.20) now follows from the following result.
Theorem 3.8. Let A : X → X be bounded and V0 –elliptic, i.e.

2
Av, v ≥ cA
1 vX for all v ∈ V0 := ker B,
where B : X → Π . For f ∈ ImVg A and g ∈ ImX B there exists a unique

solution u ∈ X of the operator equation Au = f satisfying the constraint
Bu = g.
Proof. Since g ∈ ImX B is satisfied there exists at least one ug ∈ X with

Bug = g. It remains to find u0 = u − ug ∈ V0 satisfying the operator equation
Au0 = f − Aug
which is equivalent to the variational problem
Au0 , v = f − Aug , v for all v ∈ V0 .
From the assumption f ∈ ImVg A we conclude f − Aug ∈ ImV0 A. Then there

exists at least one u0 ∈ V0 with Au0 = f − Aug . It remains to show the
uniqueness of u0 ∈ V0 . Let ū0 ∈ V0 be another solution with Aū0 = f − Aug .
From the V0 –ellipticity of A we then obtain
2
0 ≤ cA
1 u0 − ū0 X ≤ A(u0 − ū0 ), u0 − ū0 = Au0 − Aū0 , u0 − ū0 = 0
and therefore u0 = ū0 in X.

Note that ug ∈ Vg is in general not unique. However, the final solution
u = u0 + ug is unique independent of the chosen ug ∈ Vg : For ûg ∈ X with
B ûg = g there exists a unique û0 ∈ V0 satisfying A(û0 + ûg ) = f . Due to
B(ug − ûg ) = Bug − B ûg = g − g = 0 in Π
we have ug − ûg ∈ ker B = V0 . Using
A(u0 + ug ) = f, A(û0 + ûg ) = f
we obtain
A(u0 + ug − û0 − ûg ) = 0.
Obviously, u0 − û0 + (ug − ûg ) ∈ V0 , and from the V0 –ellipticity of A we
conclude
u0 − û0 + (ug − ûg ) = 0
and therefore uniqueness, u = u0 + ug = û0 + ûg .

For what follows we assume that for g ∈ ImX B there exists a ug ∈ Vg

satisfying
ug X ≤ cB gΠ (3.21)
with some positive constant cB . Then we can bound the norm of the unique
solution u ∈ Vg satisfying the variational problem (3.20) by the norms of the
given data f ∈ X and g ∈ Π .
Corollary 3.9. Let us assume the assumptions of Theorem 3.8 as well as

assumption (3.21). The solution u ∈ Vg of Au = f satisfies the estimate

1 cA
uX ≤ A f X + 1 + 2A cB gΠ .
c1 c1
Proof. Applying Theorem 3.8, the solution u of Au = f admits the represen-

tation u = u0 + ug where u0 ∈ V0 is the unique solution of the variational
problem
Au0 , v = f − Aug , v for all v ∈ V0 .
From the V0 –ellipticity of A we obtain
2
1 u0 X ≤ Au0 , u0 = f − Aug , u0 ≤ f − Aug X u0 X
cA
and therefore
1 0 1
u0 X ≤ A
f X + cA
2 ug X .
c1
Now the assertion follows from the triangle inequality and using assumption
(3.21).

3.5 Mixed Formulations
Instead of the operator equation Au = f with the constraint Bu = g we may

introduce a Lagrange multiplier p ∈ Π to formulate an extended variational
problem: Find (u, p) ∈ X × Π such that
Au, v + Bv, p = f, v

(3.22)
Bu, q = g, q
is satisfied for all (v, q) ∈ X × Π. Note that for any solution (u, p) ∈ X × Π of
the extended variational problem (3.22) we conclude that u ∈ Vg is a solution
of Au = f . The second equation in (3.22) describes just the constraint u ∈ Vg
while the first equation in (3.22) coincides with the variational formulation
to find u0 ∈ V0 when choosing as test function v ∈ V0 . It remains to ensure
the existence of the Lagrange multiplier p ∈ Π such that the first equation in
(3.22) is satisfied for all v ∈ X, see Theorem 3.11.
3.5 Mixed Formulations 53
For the Lagrange functional

1
L(v, q) := Av, v − f, v + Bv, q − g, q,
2
which is defined for (v, q) ∈ X ×Π, we first find the following characterization.
Theorem 3.10. Let A : X → X be a self–adjoint bounded and positive semi–

definite operator, i.e. Av, v ≥ 0 for all v ∈ X. Further, let B : X → Π be
bounded. (u, p) ∈ X × Π is a solution of the variational problem (3.22) iff
L(u, q) ≤ L(u, p) ≤ L(v, p) for all (v, q) ∈ X × Π. (3.23)
Proof. Let (u, p) be a solution of the variational problem (3.22). From the
first equation in (3.22) we then obtain
1
L(v, p) − L(u, p) = Av, v − f, v + Bv, p − g, p
2
1
− Au, u + f, u − Bu, p + g, p
2
1
= A(u − v), u − v + Au, v − u + B(v − u), p − f, v − u
2
1
= A(u − v), u − v ≥ 0,
2
and therefore
L(u, p) ≤ L(v, p) for all v ∈ X.
Using the second equation of (3.22) this gives
1
L(u, p) − L(u, q) = Au, u − f, u + Bu, p − g, p
2
1
− Au, u + f, u − Bu, q + g, q
2
= Bu, p − q − g, p − q = 0
and therefore
L(u, q) ≤ L(u, p) for all q ∈ Π.
For a fixed p ∈ Π we consider u ∈ X as the solution of the minimization
problem
L(u, p) ≤ L(v, p) for all v ∈ X.
Then we have for any arbitrary w ∈ X
d
L(u + tw, p)|t=0 = 0. (3.24)
dt
From
1 1
L(u + tw, p) = Au, u − f, u + Bu, p − g, p + t2 Aw, w
2 2
+ t [Au, w + Bw, p − f, w] ,
and using (3.24) we obtain the first equation of (3.22),
Au, w + Bw, p − f, w = 0 for all w ∈ X.
Now, let p ∈ Π satisfy
L(u, q) ≤ L(u, p) for all q ∈ Π.
For an arbitrary q ∈ Π we define q := p + q. Then,
0 ≤ L(u, p) − L(u, p + q)
1
= Au, u − f, u + Bu, p − g, p
2
1
− Au, u + f, u − Bu, p + q + g, p + q
2
= −Bu, q + g, q .
For q := p − q we obtain in the same way
0 ≤ L(u, p) − L(u, p − q) = Bu, q − g, q,
and therefore,
Bu, q = g, q for all q ∈ Π,
which is the second equation of (3.22).
Any solution (u, p) ∈ X × Π of the variational problem (3.22) is hence
a saddle point of the Lagrange functional L(·, ·). This is why the variational
problem (3.22) is often called a saddle point problem. The unique solvability
of (3.22) now follows from the following result.
Theorem 3.11. Let X and Π be Banach spaces and let A : X → X and

B : X → Π be bounded operators. Further, we assume that A is V0 –elliptic,
2
Av, v ≥ cA
1 vX for all v ∈ V0 = ker B ,
and that the stability condition
Bv, q
cS qΠ ≤ sup for all q ∈ Π (3.25)
0=v∈X vX
is satisfied.
For g ∈ ImX B and f ∈ ImVg A there exists a unique solution (u, p) ∈ X × Π

of the variational problem (3.22) satisfying

1 cA
2
uX ≤ A f X + 1 + A cB gΠ , (3.26)
c1 c1
and
1 cA
pΠ ≤ 1 + 2A f X + cB cA
2 gΠ . (3.27)
cS c1
Proof. Applying Theorem 3.8 we first find a unique u ∈ X satisfying
Au, v = f, v for all v ∈ V0
and
The estimate (3.26) is just the estimate of Corollary 3.9.
It remains to find p ∈ Π as the solution of the variational problem
Bv, p = f − Au, v for all v ∈ X.
First we have f − Au ∈ (ker B)0 , and using Theorem 3.6 we obtain f − Au ∈

ImΠ (B ) and therefore the solvability of the variational problem.
To prove the uniqueness of p ∈ Π we assume that there are given two
arbitrary solutions p, p̂ ∈ Π satisfying
Bv, p = f − Au, v for all v ∈ X
and
Bv, p̂ = f − Au, v for all v ∈ X.
Then,
Bv, p − p̂ = 0 for all v ∈ X.
Using the stability condition (3.25) we obtain
Bv, p − p̂
0 ≤ cS p − p̂Π ≤ sup = 0
0=v∈X vX
and therefore p = p̂ in Π.
Using again (3.25) for the unique solution p ∈ Π this gives
Bv, p f − Au, v
cS pΠ ≤ sup = sup ≤ f X + cA
2 uX
and applying (3.26) we finally obtain (3.27).

The statement of Theorem 3.11 remains valid when we assume that

A : X → X is X–elliptic, i.e.
2
Av, v ≥ cA
For an arbitrary p ∈ Π there exists a unique solution u = A−1 [f − B p] ∈ X

of the first equation of (3.22). Inserting this into the second equation of (3.22)
we obtain a variational problem to find p ∈ Π such that
BA−1 B p, q = BA−1 f − g, q (3.28)
is satisfied for all q ∈ Π. To investigate the unique solvability of the vari-

ational problem (3.28) we have to check the assumptions of Theorem 3.4
(Lax–Milgram theorem).
Lemma 3.12. Let the assumptions of Theorem 3.11 be satisfied. The operator
S := BA−1 B : Π → Π is then bounded and from the stability condition
(3.25) it follows that S is Π–elliptic,
Sq, q ≥ cS1 q2Π for all q ∈ Π. (3.29)
Proof. For q ∈ Π we have u := A−1 B q as unique solution of the variational

problem
Au, v = Bv, q for all v ∈ X.
Using the X–ellipticity of A : X → X and applying Theorem 3.4 we conclude
the existence of the unique solution u ∈ X satisfying
1 cB
2
uX = A−1 B qX ≤ B
qX ≤ qΠ .
cA
1 cA
1
From this we obtain

2
[cB
2 ]
SqΠ = BA−1 B qΠ = BuΠ ≤ cB
2 uX ≤ qΠ
cA
1
for all q ∈ Π and therefore the boundedness of S : Π → Π . Further,
Sq, q = BA−1 B q, q = Bu, q = Au, u ≥ cA 2

1 uX .
On the other hand, the stability condition (3.25) gives

Bv, q Au, v
cS qΠ ≤ sup = sup ≤ cA
2 uX
A 2
1 [cS /c2 ] .
and therefore the ellipticity estimate (3.29) with cS1 = cA
From Lemma 3.12 we see that (3.28) is an elliptic variational problem to
find p ∈ Π. Hence we obtain the unique solvability of (3.28) when applying
Theorem 3.4. Moreover, for the solution of the variational problem (3.22) we
obtain the following result.
3.6 Coercive Operators 57
Theorem 3.13. Let X and Π be Banach spaces and let A : X → X and

B : X → Π be bounded operators. We assume that A is X–elliptic, and that
the stability condition (3.25) is satisfied. For f ∈ X and g ∈ Π there exists
the unique solution (u, p) ∈ X × Π of the variational problem (3.22) satisfying

1 −1 1 cB 2
pΠ ≤ S BA f − gΠ ≤ S f X + gΠ
(3.30)
c1 c1 cA 1
and 2

1 [cB
2 ] cB
2
uX ≤ 1+ f X + gΠ . (3.31)
cA
1
A S
c1 c1 c1 cS1
A
Proof. The application of Theorem 3.4 (Lax–Milgram lemma) gives the

unique solvability of the variational problem (3.28) as well as the estimate
(3.30). For a known p ∈ X we find u ∈ X as the unique solution of the
variational problem
Au, v = f − B p, v for all v ∈ X.
From the X–ellipticity of A we obtain

2
1 uX ≤ Au, u = f − B p, u ≤ f − B pX uX
cA
and therefore
1 cB
2
uX ≤ f X + pΠ .
cA
1 cA
1
Applying (3.30) this gives the estimate (3.31).

3.6 Coercive Operators

Since the ellipticity assumption (3.12) is too restrictive for some applications
we now consider the more general case of coercive operators. An operator
A : X → X is called coercive if there exists a compact operator C : X → X
such that there holds a Gårdings inequality, i.e.
2
(A + C)v, v ≥ cA
1 vX for all v ∈ X. (3.32)
An operator C : X → Y is said to be compact if the image of the unit sphere

of X is relatively compact in Y . Note that the product of a compact operator
with a bounded linear operator is compact. Applying the Riesz–Schauder
theory, see for example [163], we can state the following result.
Theorem 3.14 (Fredholm alternative). Let K : X → X be a compact

operator. Either the homogeneous equation
(I − K)u = 0
has a non–trivial solution u ∈ X or the inhomogeneous equation
(I − K)u = g
has, for every given g ∈ X, a uniquely determined solution u ∈ X satisfying
uX ≤ c gX .
Based on Fredholm’s alternative we can derive a result on the solvability

of operator equations Au = f when A is assumed to be coercive.
Theorem 3.15. Let A : X → X be a bounded coercive linear operator and

let A be injective, i.e., from Au = 0 it follows that u = 0. Then there exists
the unique solution u ∈ X of the operator equation Au = f satisfying
uX ≤ c f X .
Proof. The linear operator D = A + C : X → X is bounded and, due to

assumption (3.32), X–elliptic. Applying the Lax–Milgram lemma (Theorem
3.4) this gives the inverse operator D−1 : X → X. Hence, instead of the
operator equation Au = f we consider the equivalent equation
Bu = D−1 Au = D−1 f (3.33)
with the bounded operator
B = D−1 A = D−1 (D − C) = I − D−1 C : X → X.
Since the operator D−1 C : X → X is compact we can apply Theorem 3.14

to investigate the unique solvability of the operator equation (3.33). Since
A is assumed to be injective the homogeneous equation D−1 Au = 0 has
only the trivial solution. Hence there exists a unique solution u ∈ X of the
inhomogeneous equation Bu = D−1 f satisfying
uX ≤ c D−1 f X ≤
c f X .

4
Variational Formulations
of Boundary Value Problems
In this chapter we describe and analyze variational methods for second order
elliptic boundary value problems as given in Chapter 1. To establish the unique
solvability of the associated variational formulations we will use the methods
which were given in the previous Chapter 3. The weak formulation of boundary
value problem is the basis to introduce finite element methods. Moreover,
from these results we can also derive mapping properties of boundary integral
operators (cf. Chapter 6) as used in boundary element methods.
4.1 Potential Equation

Let us consider the scalar partial differential operator (1.1),

d
∂ ∂
(Lu)(x) = − aji (x) u(x) for x ∈ Ω ⊂ Rd , (4.1)
i,j=1
∂xj ∂xi
the trace operator (1.3),
γ0int u(x) = lim u(

x) for x ∈ Γ = ∂Ω,
x→x∈Γ
Ω
and the associated conormal derivative (1.7),

d
∂
γ1int u(x) = lim nj (x)aji (
x) u(
x) for x ∈ Γ = ∂Ω. (4.2)
x→x∈Γ
Ω
i,j=1
∂
xi
Note that Green’s first formula (1.5),

a(u, v) = (Lu)(x)v(x)dx + γ1int u(x)γ0int v(x)dsx ,
Ω Γ
60 4 Variational Formulations of Boundary Value Problems
−1 (Ω) and v ∈ H 1 (Ω), i.e. we have

remains valid for u ∈ H 1 (Ω) with Lu ∈ H
a(u, v) = Lu, vΩ + γ1int u, γ0int vΓ (4.3)
where a(·, ·) is the symmetric bilinear form as defined in (1.6),

d
∂ ∂
a(u, v) = aji (x) u(x) v(x) dx. (4.4)
i,j=1 Ω
∂xi ∂xj
Lemma 4.1. Assume that aij ∈ L∞ (Ω) for i, j = 1, . . . , d with
aL∞ (Ω) := max sup |aij (x)|. (4.5)

i,j=1,...,d x∈Ω
The bilinear form a(·, ·) : H 1 (Ω) × H 1 (Ω) → R is bounded satisfying
|a(u, v)| ≤ cA
2 |u|H 1 (Ω) |v|H 1 (Ω) for all u, v ∈ H 1 (Ω) (4.6)
2 := d aL∞ (Ω) .
with cA
Proof. Using (4.5) we first have

d ∂ ∂
|a(u, v)| = aji (x) u(x) v(x)dx
i,j=1 ∂xi ∂xj
Ω

d
∂ d
∂
≤ aL∞ (Ω)
∂xi u(x) ∂xj v(x) dx.
Ω i=1 j=1
Applying the Cauchy–Schwarz inequality twice we then obtain

⎛ ⎞1/2
#
d $2
∂
|a(u, v)| ≤ aL∞ (Ω) ⎝ ⎠
∂xi u(x) dx
Ω i=1
⎛ ⎡ ⎤ ⎞1/2
d 2
⎜ ∂ ⎟
⎝ ⎣ ⎦
∂xj v(x) dx⎠
Ω j=1
⎛ ⎞1/2
d

2
∂
≤ aL∞ (Ω) ⎝ d ⎠
∂xi u(x) dx
Ω i=1
⎛ ⎞1/2
d
2
∂
⎝ d ⎠
∂xj v(x) dx
Ω j=1
= d aL∞ (Ω) ∇uL2 (Ω) ∇vL2 (Ω) .

From (4.6) we further get the estimate
|a(u, v)| ≤ cA
2 uH 1 (Ω) vH 1 (Ω) for all u, v ∈ H 1 (Ω). (4.7)
Lemma 4.2. Let L be a uniform elliptic partial differential operator as given

in (4.1). For the bilinear form (4.4) we then have
a(v, v) ≥ λ0 |v|2H 1 (Ω) for all v ∈ H 1 (Ω) (4.8)
where λ0 is the positive constant of the uniform ellipticity estimate (1.2).

∂
Proof. By using wi (x) := v(x) for i = 1, . . . , d we have
∂xi

a(v, v) = (A(x)w(x), w(x)) dx
Ω

≥ λ0 (w(x), w(x)) dx = λ0 ∇v2L2 (Ω) .

Ω
4.1.1 Dirichlet Boundary Value Problem
We start to consider the Dirichlet boundary value problem (1.10) and (1.11),
(Lu)(x) = f (x) for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ. (4.9)
The manifold to be used in the weak formulation is defined as

! "
Vg := v ∈ H 1 (Ω) : γ0int v(x) = g(x) for x ∈ Γ , V0 = H01 (Ω).
The variational formulation of the Dirichlet boundary value problem (4.9)

then follows from Green’s first formula (4.3): Find u ∈ Vg such that
a(u, v) = f, vΩ (4.10)
is satisfied for all v ∈ V0 . Since the Dirichlet boundary condition is explic-

itly incorporated as a side condition in the manifold Vg , we call boundary
conditions of Dirichlet type also essential boundary conditions.
The variational problem (4.10) corresponds to the abstract formulation
(3.20). Hence we can apply Theorem 3.8 and Corollary 3.9 to establish the
unique solvability of the variational problem (4.10).
Theorem 4.3. For f ∈ H −1 (Ω) and g ∈ H 1/2 (Γ ) there exists a unique solu-
tion u ∈ H 1 (Ω) of the variational problem (4.10) satisfying

1 cA
uH 1 (Ω) ≤ A f H −1 (Ω) + 1 + 2A cIT gH 1/2 (Γ ) . (4.11)
c1 c1
Proof. For any given Dirichlet datum g ∈ H 1/2 (Γ ) we find, by applying the
inverse trace theorem (Theorem 2.22), a bounded extension ug ∈ H 1 (Ω) sat-
isfying γ0int ug = g and
ug H 1 (Ω) ≤ cIT gH 1/2 (Γ ) .
It remains to find u0 := u − ug ∈ V0 as the solution of the variational problem
a(u0 , v) = f, vΩ − a(ug , v) for all v ∈ V0 . (4.12)
Recall that
⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
vW21 (Ω),Γ := ⎣ γ0 v(x) dsx ⎦ + ∇vL2 (Ω)
int 2
⎪
⎩ ⎪
⎭
Γ
defines an equivalent norm in H 1 (Ω) (cf. Example 2.7). For v ∈ V0 = H01 (Ω)
we then find from Lemma 4.2
a(v, v) ≥ λ0 |v|2H 1 (Ω) = λ0 v2W 1 (Ω),Γ ≥ cA 2

1 vH 1 (Ω) . (4.13)
2
Therefore, all assumptions of Theorem 3.4 (Lax–Milgram lemma) are satisfied.

Hence we conclude the unique solvability of the variational problem (4.12).
For the unique solution u0 ∈ V0 of the variational problem (4.12) we have,
since the bilinear form a(·, ·) is V0 –elliptic and bounded,
2
1 u0 H 1 (Ω) ≤ a(u0 , u0 ) = f, u0 Ω − a(ug , u0 )
cA
0 1
≤ f H −1 (Ω) + cA2 ug H 1 (Ω) u0 H 1 (Ω) ,
from which we finally get the estimate (4.11).

The unique solution u ∈ Vg of the variational problem (4.10) is also
denoted as weak solution of the Dirichlet boundary value problem (4.9).
For f ∈ H −1 (Ω) we can determine the associated conormal derivative
int −1/2
γ1 u ∈ H (Γ ) as the solution of the variational problem
γ1int u, zΓ = a(u, Ez) − f, EzΩ (4.14)
for all z ∈ H 1/2 (Γ ). In (4.14), E : H 1/2 (Γ ) → H 1 (Ω) is the bounded exten-

sion operator as defined by the inverse trace theorem (Theorem 2.22). The
unique solvability of the variational formulation (4.14) follows when applying
Theorem 3.7. Hence we need to assume the stability condition
w, zΓ
wH −1/2 (Γ ) = sup for all w ∈ H −1/2 (Γ ). (4.15)
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
Lemma 4.4. Let u ∈ H 1 (Ω) be the unique solution of the Dirichlet boundary
value problem (4.10) when assuming g ∈ H 1/2 (Γ ) and f ∈ H −1 (Ω). For the
int −1/2
associated conormal derivative γ1 u ∈ H (Γ ) we then have
! "
γ1int uH −1/2 (Γ ) ≤ cIT f H
-−1 (Ω) + c2 |u|H 1 (Ω) .
A
(4.16)
Proof. Using the stability condition (4.15) and the variational formulation
(4.14) we find from the boundedness of the bilinear form a(·, ·) and by applying
the inverse trace theorem
|γ1int u, zΓ |
γ1int uH −1/2 (Γ ) = sup
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
|a(u, Ez) − f, EzΩ |

= sup
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
! " EzH 1 (Ω)
≤ cA2 |u|H 1 (Ω) + f H-−1 (Ω) sup
0=z∈H 1/2 (Γ ) zH 1/2 (Γ )
! "
≤ cIT f H
-−1 (Ω) + c2 |u|H 1 (Ω) .
A

In particular for the solution u of the Dirichlet boundary value problem

with a homogeneous partial differential equation, i.e. f ≡ 0, we obtain the
following result which is essential for the analysis of boundary integral oper-
ators.
Corollary 4.5. Let u ∈ H 1 (Ω) be the weak solution of the Dirichlet boundary
value problem
(Lu)(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ
where L is a uniform elliptic partial differential operator of second order.

Then,
a(u, u) ≥ c γ1int u2H −1/2 (Γ ) . (4.17)
Proof. By setting f ≡ 0 the estimate (4.16) first gives
γ1int u2H −1/2 (Γ ) ≤ [cIT cA 2 2

2 ] |u|H 1 (Ω) .
The assertion now follows from the semi–ellipticity (4.8) of the bilinear form
a(·, ·).

When Ω is a Lipschitz domain we can formulate stronger assumptions on
the given data f and g to establish higher regularity results for the solution
u of the Dirichlet boundary value problem and for the associated conormal
derivative γ1int u.
Theorem 4.6. [106, Theorem 1.1, p. 249] Let Ω ⊂ Rd be a bounded Lipschitz

domain with boundary Γ = ∂Ω. Let u ∈ H 1 (Ω) be the weak solution of the
Dirichlet boundary value problem
(Lu)(x) = f (x) for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ.
If f ∈ L2 (Ω) and g ∈ H 1 (Γ ) are satisfied, then we have u ∈ H 3/2 (Ω) with

uH 3/2 (Ω) ≤ c1 f L2 (Ω) + gH 1 (Γ )
as well as γ1int u ∈ L2 (Γ ) satisfying

γ1int uL2 (Γ ) ≤ c2 f L2 (Ω) + gH 1 (Γ ) .
When formulating stronger assumptions both on the domain Ω and on
the given data f and g we can establish even higher regularity results for the
solution u of the Dirichlet boundary value problem (4.9). Let the boundary
Γ = ∂Ω be either smooth or piecewise smooth, but assume that Ω is convex,
and let f ∈ L2 (Ω). If g = γ0int ug is the trace of a function ug ∈ H 2 (Ω), then
we have u ∈ H 2 (Ω). For more general results on the regularity of solutions of
boundary value problems we refer, for example, to [66].
4.1.2 Lagrange Multiplier Methods

In what follows we will consider a saddle point variational formulation which
is equivalent to the variational problem (4.10). The Dirichlet boundary con-
ditions are now formulated as side conditions, and the associated conor-
mal derivative corresponds to the Lagrange multiplier [7, 24]. Starting from
Green’s first formula (4.3) we obtain by introducing the Lagrange multi-
plier λ := γ1int u ∈ H −1/2 (Γ ) the following saddle point problem: Find
(u, λ) ∈ H 1 (Ω) × H −1/2 (Γ ) such that
a(u, v) − b(v, λ) = f, vΩ
(4.18)
b(u, µ) = g, µΓ
is satisfied for all (v, µ) ∈ H 1 (Ω) × H −1/2 (Γ ). Here we have used the bilinear
form
b(v, µ) := γ0int v, µΓ for (v, µ) ∈ H 1 (Ω) × H −1/2 (Γ ).
To investigate the unique solvability of the saddle point problem (4.18) we
will apply Theorem 3.11. Obviously,
! "
ker B := v ∈ H 1 (Ω) : γ0int v, µΓ = 0 for all µ ∈ H −1/2 (Γ ) = H01 (Ω).
Hence, due to (4.13), we have the ker B–ellipticity of the bilinear form a(·, ·).
It remains to establish the stability condition
γ0int v, µΓ
cS µH −1/2 (Γ ) ≤ sup for all µ ∈ H −1/2 (Γ ) . (4.19)
0=v∈H 1 (Ω) vH 1 (Ω)
Lemma 4.7. The stability condition (4.19) is satisfied for all µ ∈ H −1/2 (Γ ).
Proof. Let an arbitrary µ ∈ H −1/2 (Γ ) be given. Applying Theorem 3.3 (Riesz

Representation Theorem) we find a uniquely determined uµ ∈ H 1/2 (Γ ) satis-
fying
uµ , vH 1/2 (Γ ) = µ, vΓ for all v ∈ H 1/2 (Γ )
and
uµ H 1/2 (Γ ) = µH −1/2 (Γ ) .
Using the inverse trace theorem (Theorem 2.22) there exists an extension
Euµ ∈ H 1 (Ω) with
Euµ H 1 (Ω) ≤ cIT uµ H 1/2 (Γ ) .
For v = Euµ ∈ H 1 (Ω) we then have
v, µΓ uµ , µΓ uµ , uµ H 1/2 (Γ )

= =
vH 1 (Ω) Euµ H 1 (Ω) Euµ H 1 (Ω)
1 1
≥ uµ H 1/2 (Γ ) = µH −1/2 (Γ )
cIT cIT
and therefore the stability condition (4.19) is satisfied.

Hence we can conclude the unique solvability of the saddle point problem
(4.18) due to Theorem 3.11.
Recall that the bilinear form a(·, ·) in the saddle point formulation (4.18)
is only H01 (Ω)–elliptic. However, the saddle point problem (4.18) can be re-
formulated to obtain a formulation where the modified bilinear form a(·, ·) is
now H 1 (Ω)–elliptic. Since the Lagrange multiplier λ := γ1int u ∈ H −1/2 (Γ )
describes the conormal derivative of the solution u, using the orthogonality
relation (1.15) we have

f (x)dx + λ(x)dsx = 0 . (4.20)
Ω Γ
On the other hand, with the Dirichlet boundary condition γ0int u = g we also
have
γ0int u(x)dsx = g(x)dsx . (4.21)
Γ Γ
Hence we can reformulate the saddle point problem (4.18) to find (u, λ) ∈
H 1 (Ω) × H −1/2 (Γ ) such that

γ0int u(x)dsx γ0int v(x)dsx + a(u, v) − b(v, λ) (4.22)
Γ Γ

= f, vΩ + g(x)dsx γ0int v(x)dsx
Γ Γ

b(u, µ) + λ(x)dsx µ(x)dsx = g, µΓ − f (x)dx µ(x)dsx (4.23)
Γ Γ Ω Γ
1 −1/2
is satisfied for all (v, µ) ∈ H (Ω) × H (Γ ).
The modified saddle point problem (4.22) and (4.23) is uniquely solv-
able, and the solution is also the unique solution of the original saddle point
problem (4.18), i.e. the saddle point formulations (4.22)–(4.23) and (4.18) are
equivalent.
Theorem 4.8. The modified saddle point problem (4.22) and (4.23) has a
unique solution (u, λ) ∈ H 1 (Ω) × H −1/2 (Γ ), which is also the unique solution
of the saddle point formulation (4.18).
Proof. The extended bilinear form

a(u, v) := γ0int u(x)dsx γ0int v(x)dsx + a(u, v)
Γ Γ
1
is bounded for all u, v ∈ H (Ω). Using Lemma 4.2 and Example 2.7 we find
⎡ ⎤2

a(v, v) = ⎣ γ0int v(x)dsx ⎦ + a(v, v) ≥ min{1, λ0 }v2 1
≥ cA 2
1 vH 1 (Ω) W2 (Ω),Γ
Γ
for all v ∈ H 1 (Ω) and therefore the H 1 (Ω)–ellipticity of the extended bilinear
form a(·, ·). Applying Theorem 3.11 we obtain as in Theorem 3.13 the unique
solvability of the saddle point problem (4.22) and (4.23). In particular for
(v, µ) ≡ (1, 1) we have

|Γ | γ0int u(x)dsx − λ(x)dsx = f (x)dx + |Γ | g(x)dsx ,
Γ Γ Ω Γ

γ0int u(x)dsx + |Γ | λ(x)dsx = g(x)dsx − |Γ | f (x)dx.
Γ Γ Γ Ω
Multiplying the first equation with |Γ | > 0 and adding the result to the second
equation this gives

2 int 2
(1 + |Γ | ) γ0 u(x)dsx = (1 + |Γ | ) g(x)dsx
Γ Γ
and therefore (4.21). Then we immediately get also (4.20), i.e. (u, λ) is also a
solution of the saddle point problem (4.18).

4.1.3 Neumann Boundary Value Problem
In addition to the Dirichlet boundary value problem (4.9) we now consider

the Neumann boundary value problem (1.10) and (1.12),
(Lu)(x) = f (x) for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ. (4.24)
Hereby we have to assume the solvability condition (1.17),

f (x)dx + g(x)dsx = 0. (4.25)
Ω Γ
Moreover, the solution of the Neumann boundary value problem (4.24) is only
unique up to an additive constant. To fix this constant, we formulate a suitable
scaling condition. For this we define
⎧ ⎫
⎨ ⎬
H∗1 (Ω) := v ∈ H 1 (Ω) : v(x)dx = 0 .
⎩ ⎭
Ω
Using Green’s first formula (4.3) we obtain the variational formulation of the
Neumann boundary value problem (4.24) to find u ∈ H∗1 (Ω) such that
a(u, v) = f, vΩ + g, γ0int vΓ (4.26)
is satisfied for all v ∈ H∗1 (Ω).
Theorem 4.9. Let f ∈ H −1 (Ω) and g ∈ H −1/2 (Γ ) be given satisfying the

solvability condition (4.25). Then there exists a unique solution u ∈ H∗1 (Ω) of
the variational problem (4.26) satisfying
1 ! "
uH 1 (Ω) ≤ f -−1
H (Ω)
+ cT gH −1/2 (Γ ) .

cA
1
Proof. Recall that

⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
vW21 (Ω),Ω := ⎣ v(x) dx⎦ + ∇vL2 (Ω)
2
⎪
⎩ ⎪
⎭
Ω
defines an equivalent norm in H 1 (Ω) (cf. Example 2.7). Using Lemma 4.2 we
then have
a(v, v) ≥ λ0 ∇v2L2 (Ω) = λ0 v2W 1 (Ω),Ω ≥ 2

1 vH 1 (Ω)
cA (4.27)
2
for all v ∈ H∗1 (Ω) and therefore the H∗1 (Ω)–ellipticity of the bilinear form
a(·, ·) follows. The unique solvability of the variational problem (4.26) we
now conclude from Theorem 3.4 (Lax–Milgram lemma). Using the H∗1 (Ω)–
ellipticity of the bilinear form a(·, ·) we further have
2 int
1 uH 1 (Ω) ≤ a(u, u) = f, uΩ + g, γ0 uΓ
cA
≤ f H int
-−1 (Ω) uH 1 (Ω) + gH −1/2 (Γ ) γ0 uH 1/2 (Γ ) .
Applying the trace theorem (Theorem 2.21) gives the assertion.

In what follows we will consider a saddle point formulation which is equiv-
alent to the variational problem (4.26). The scaling condition to define the
trial space H∗1 (Ω) is now formulated as a side condition. By using a scalar
Lagrange multiplier we obtain the following variational problem (cf. Section
3.5) to find u ∈ H 1 (Ω) and λ ∈ R such that

a(u, v) + λ v(x)dx = f, vΩ + g, γ0int vΓ
Ω (4.28)
u(x)dx =0
Ω
is satisfied for all v ∈ H 1 (Ω). To establish the unique solvability of the saddle
point problem (4.28) we have to investigate the assumptions of Theorem 3.11.
The bilinear form

b(v, µ) := µ v(x)dx for all v ∈ H 1 (Ω), µ ∈ R
Ω
is bounded, and we have ker B = H∗1 (Ω). Hence we obtain the ker B–ellipticity
of the bilinear form a(·, ·) from the ellipticity estimate (4.27). It remains to
prove the stability condition
b(v, µ)
cS |µ| ≤ sup for all µ ∈ R. (4.29)
0=v∈H 1 (Ω) vH 1 (Ω)
∗ 1
For an arbitrary given µ ∈ R we define . v := µ ∈ H (Ω) to obtain the
stability estimate (4.29) with cS = 1/ |Ω|. By applying Theorem 3.11 we
now conclude the unique solvability of the saddle point problem (4.28).
Choosing in (4.28) the test function v ≡ 1 we obtain for the Lagrange
parameter λ from the solvability condition (4.25)
λ = 0.
Instead of (4.28) we may now consider an equivalent saddle point formulation
to find (u, λ) ∈ H 1 (Ω) × R such that

a(u, v) + λ v(x)dx = f, vΩ + g, γ0int vΓ
Ω (4.30)
u(x)dx − λ =0
Ω
is satisfied for all v ∈ H 1 (Ω). Using the second equation we can eliminate the
scalar Lagrange multiplier λ ∈ R to obtain a modified variational problem to
find u ∈ H 1 (Ω) such that

a(u, v) + u(x)dx v(x)dx = f, vΩ + g, γ0int vΓ (4.31)
Ω Ω
is satisfied for all v ∈ H 1 (Ω).
Theorem 4.10. For any f ∈ H −1 (Ω) and for any g ∈ H −1/2 (Γ ) there is a
1
unique solution u ∈ H (Ω) of the modified variational problem (4.31).
If f ∈ H −1 (Ω) and g ∈ H −1/2 (Γ ) satisfy the solvability condition (4.25),
then we have u ∈ H∗1 (Ω), i.e. the modified variational problem (4.31) and the
saddle point formulation (4.28) are equivalent.
Proof. The modified bilinear form

a(u, v) := a(u, v) + u(x)dx v(x)dx
Ω Ω
is H 1 (Ω)–elliptic, i.e for all v ∈ H 1 (Ω) we have

⎡ ⎤2

a(v, v) ≥ λ0 ∇v2L2 (Ω) + ⎣
v(x)dx⎦
Ω
≥ min{λ0 , 1} v2W 1 (Ω),Ω ≥ ĉA 2

1 vH 1 (Ω) .
2
Hence we conclude the unique solvability of the modified variational problem

(4.31) due the Theorem 3.4 (Lax–Milgram lemma) for arbitrary given data
f ∈H −1 (Ω) and g ∈ H −1/2 (Γ ).
Choosing as test function v ≡ 1 we get, when assuming the solvability
condition (4.25),

|Ω| u(x)dx = f, 1Ω + g, 1Γ = 0
Ω
and therefore u ∈ H∗1 (Ω). The solution of the modified variational problem
(4.31) is therefore also a solution of the saddle point formulation (4.28), i.e.
both formulations are equivalent.
Since the solution of the Neumann boundary value problem (4.24) is not
unique, we can add an arbitrary constant α ∈ R to the solution u ∈ H∗1 (Ω)
to obtain the general solution u := u + α ∈ H 1 (Ω).
4.1.4 Mixed Boundary Value Problem
We now consider the boundary value problem (1.10)–(1.12) with boundary

conditions of mixed type,
(Lu)(x) = f (x) for x ∈ Ω,

γ0int u(x) = gD (x) for x ∈ ΓD ,
γ1int u(x) = gN (x) for x ∈ ΓN .
We assume Γ = Γ D ∪ Γ N as well as meas(ΓD ) > 0. The associated variational

problem again follows from Green’s first formula (4.3) to find u ∈ H 1 (Ω) with
γ0int u(x) = gD (x) for x ∈ ΓD such that
a(u, v) = f, vΩ + gN , γ0int vΓN (4.32)
is satisfied for all v ∈ H01 (Ω, ΓD ) where

! "
H01 (Ω, ΓD ) := v ∈ H 1 (Ω) : γ0int v(x) = 0 for x ∈ ΓD .
The unique solvability of the variational problem (4.32) is a consequence of

the following theorem.
Theorem 4.11. Let f ∈ H −1 (Ω), gD ∈ H 1/2 (ΓD ) and gN ∈ H −1/2 (ΓN )
be given. Then there exists a unique solution u ∈ H 1 (Ω) of the variational
problem (4.32) satisfying
2 3
uH 1 (Ω) ≤ c f H-−1 (Ω) + gD H 1/2 (ΓD ) + gN H −1/2 (ΓN ) . (4.33)
Proof. For gD ∈ H 1/2 (ΓD ) we first find a bounded extension gD ∈ H 1/2 (Γ )
satisfying
gD H 1/2 (Γ ) ≤ c gD H 1/2 (ΓD ) .

Applying the inverse trace theorem (Theorem 2.22) there exists a second ex-
tension ugD ∈ H 1 (Ω) with γ0int ugD = gD and satisfying
ugD H 1 (Ω) ≤ cIT

gD H 1/2 (Γ ) .
It remains to find u0 ∈ H01 (Ω, ΓD ) as the unique solution of the variational

formulation
a(u0 , v) = f, vΩ + gN , γ0int vΓN − a(ugD , v)
for all v ∈ H01 (Ω, ΓD ). As in Example 2.7 we can define

⎧⎡ ⎤2 ⎫1/2
⎪
⎨ ⎪
⎬
vW21 (Ω),ΓD := ⎣ γ0int v(x) dsx ⎦ + ∇v2L2 (Ω)
⎪
⎩ ⎪
⎭
ΓD
which is an equivalent norm in H 1 (Ω). Using Lemma 4.2 we find
a(v, v) ≥ λ0 ∇v2L2 (Ω) = λ0 v2W 1 (Ω),ΓD ≥ cA 2

1 vW 1 (Ω)
2 2
for all v ∈ H01 (Ω, ΓD ).

Hence all assumptions of Theorem 3.4 (Lax–Milgram
lemma) are satisfied, and therefore, the unique solvability of the variational
problem (4.32) follows.
For the unique solution u0 ∈ H01 (Ω, ΓD ) of the variational problem (4.32)
we then obtain
2
1 u0 H 1 (Ω) ≤ a(u0 , u0 )
cA
= f, u0 Ω + gN , γ0int u0 ΓN − a(ugD , u0 )

2 3
≤ f H-−1 (Ω) + c2 ug
A
D H 1 (Ω) u0 H 1 (Ω)
+gN H −1/2 (ΓN ) γ0int u0 H

-1/2 (ΓN ) ,
from which we conclude the estimate (4.33).

4.1.5 Robin Boundary Value Problems
We finally consider the boundary value problem (1.10) and (1.13) with bound-
ary conditions of Robin type,
(Lu)(x) = f (x) for x ∈ Ω, γ1int u(x) + κ(x)γ0int u(x) = g(x) for x ∈ Γ.
The associated variational formulation is again a direct consequence of Green’s

first formula (4.3) to find u ∈ H 1 (Ω) such that

a(u, v) + κ(x)γ0int u(x)γ0int v(x)dsx = f, vΩ + g, γ0int vΓ (4.34)
Γ
is satisfied for all v ∈ H 1 (Ω).
Theorem 4.12. Let f ∈ H −1 (Ω) and g ∈ H −1/2 (Γ ) be given. Assume that

κ(x) ≥ κ0 > 0 holds for all x ∈ Γ . Then there exists a unique solution
u ∈ H 1 (Ω) of the variational problem (4.34) satisfying
2 3
uH 1 (Ω) ≤ c f H-−1 (Ω) + gH −1/2 (Γ ) . (4.35)
Proof. As in Example 2.7 we can define

! "1/2
vW21 (Ω),Γ := γ0int v2L2 (Γ ) + ∇v2L2 (Ω)
which is an equivalent norm in H 1 (Ω). Applying Lemma 4.2 and by using

κ(x) ≥ κ0 > 0 for x ∈ Γ we obtain

a(v, v) + κ(x)[γ0int v(x)]2 dsx ≥ λ0 ∇v2L2 (Ω) + κ0 γ0int v2L2 (Γ )
Γ
≥ min{λ0 , κ0 } v2H 1 (Ω),Γ ≥ cA 2
1 vH 1 (Ω) .
Hence we can apply Theorem 3.4 (Lax–Milgram lemma) to conclude the

unique solvability of the variational problem (4.34). The estimate (4.35) then
follows as in the proof of Theorem 4.11.

Next we consider the system of linear elasticity,
d
∂
Li u(x) = − σij (u, x) for x ∈ Ω ⊂ Rd , i = 1, . . . , d.
j=1
∂x j
Inserting Hooke’s law (1.19) this is equivalent to
Lu(x) = −µ∆u(x) − (λ + µ)grad div u(x) for x ∈ Ω ⊂ Rd
with the Lamé constants

Eν E
λ = , µ =
(1 + ν)(1 − 2ν) 2(1 + ν)
where we assume E > 0 and ν ∈ (0, 1/2). Using the associated conormal
derivative (1.23) and the bilinear form (1.26),

d
a(u, v) = 2 µ eij (u, x)eij (v, x)dx + λ div u(x) div v(x)dx
Ω i,j=1 Ω

d
= σij (u, x)eij (v, x)dx,
Ω i,j=1
we can write Betti’s first formula as (1.21),
a(u, v) = Lu, vΩ + γ1int u, γ0int vΓ .
First we show that the bilinear form a(·, ·) is bounded.
Lemma 4.13. The bilinear form (1.26) is bounded, i.e.

2E
|a(u, v)| ≤ |u| 1 d |v|[H 1 (Ω)]d (4.36)
1 − 2ν [H (Ω)]
for all u, v ∈ [H 1 (Ω)]d .
Proof. In the case d = 3 we can write Hooke’s law (1.19) as

⎛ ⎞ ⎛ ⎞⎛ ⎞
σ11 1−ν ν ν e11
⎜ σ22 ⎟ ⎜ ν 1−ν ν ⎟ ⎜ e22 ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ σ33 ⎟ E ⎜ ν ν 1−ν ⎟ ⎜ e33 ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎜ σ12 ⎟ = (1 + ν)(1 − 2ν) ⎜ 1 − 2ν ⎟ ⎜ e12 ⎟
⎜ ⎟ ⎜ ⎟⎜ ⎟
⎝ σ13 ⎠ ⎝ 1 − 2ν ⎠ ⎝ e13 ⎠
σ23 1 − 2ν e23
or in short,
E
σ = C e.
(1 + ν)(1 − 2ν)
Due to the symmetries σij (u, x) = σji (u, x) and eij (v, x) = eji (v, x) we have

E
a(u, v) = (DCe(u, x), e(v, x)) dx
(1 + ν)(1 − 2ν)
Ω
with the diagonal matrix D = diag(1, 1, 1, 2, 2, 2). The eigenvalues of the ma-
trix DC ∈ R6×6 are
λ1 (DC) = 1 + ν, λ2,3 (DC) = 1 − 2ν, λ4,5,6 (DC) = 2(1 − 2ν).
Hence, by applying the Cauchy–Schwarz inequality,

E

|a(u, v)| = (DCe(u, x), e(v, x)) dx
(1 + ν)(1 − 2ν)
Ω

E
≤ DCe(u, x)2 e(v, x)2 dx
(1 + ν)(1 − 2ν)
Ω

E
≤ max{(1 + ν), 2(1 − 2ν)} e(v, x)2 e(v, x)2 dx
(1 + ν)(1 − 2ν)
Ω
⎛ ⎞1/2 ⎛ ⎞1/2

2E ⎝ e(u, x)22 dx⎠ ⎝ e(v, x)22 dx⎠ .
≤
1 − 2ν
Ω Ω
Using

d d 2
1 ∂
d d
∂
e(v, x)22 = 2
[eij (u, x)] = uj (x) + ui (x)
i=1 j=i
4 i=1 j=i ∂xi ∂xj
2 2 + d 2
1
d
∂ ∂ ∂
≤ ui (x) + uj (x) = ui (x)
2 i,j=1 ∂xj ∂xi i,j=1
∂xj
we obtain
d 2
∂
e(u, x)22 dx ≤ ui (x) dx = |u|2[H 1 (Ω)]d
i,j=1
∂x j
Ω Ω
and therefore the estimate (4.36). In the case d = 2 the assertion follows in
the same way.
The proof of the [H01 (Ω)]d –ellipticity of the bilinear form a(·, ·) requires
several steps.
Lemma 4.14. For v ∈ [H 1 (Ω)]d we have

d
E
a(v, v) ≥ [eij (v, x)]2 dx . (4.37)
1+ν
Ω i,j=1
Proof. The assertion follows from the representation (1.26), i.e.

d
2
a(v, v) = 2µ [eij (v, x)] dx + λ [div v(x)]2 dx
Ω i,j=1 Ω

d
d
E
≥ 2µ [eij (v, x)]2 dx = [eij (v, x)]2 dx.

1+ν
Ω i,j=1 Ω i,j=1
Next we can formulate Korn’s first inequality for v ∈ [H01 (Ω)]d .
Lemma 4.15 (Korn’s First Inequality). For v ∈ [H01 (Ω)]d we have

d
1 2
[eij (v, x)]2 dx ≥ |v| 1 d. (4.38)
2 [H (Ω)]
Ω i,j=1
Proof. For ϕ ∈ C0∞ (Ω) we first have

d d 2
1 ∂ ∂
[eij (ϕ, x)]2 dx = ϕi (x) + ϕj (x) dx
4 ∂xj ∂xi
Ω i,j=1 Ω i,j=1
d 2 d
1 ∂ 1 ∂ ∂
= ϕi (x) dx + ϕi (x) ϕj (x)dx.
2 i,j=1 ∂xj 2 i,j=1 ∂xj ∂xi
Ω Ω
Applying integration by parts twice, this gives

∂ ∂ ∂ ∂
ϕi (x) ϕj (x)dx = ϕi (x) ϕj (x)dx
∂xj ∂xi ∂xi ∂xj
Ω Ω
and therefore
d
d
∂ ∂ ∂ ∂
ϕi (x) ϕj (x)dx = ϕi (x) ϕj (x)dx
i,j=1 Ω
∂xj ∂xi i,j=1
∂xi ∂xj
Ω
#d
$2
∂
= ϕi (x) dx ≥ 0.
i=1
∂xi
Ω
Hence we have
d 2
1
d
2 ∂ 1
[eij (ϕ, x)] dx ≥ ϕi (x) dx = |ϕ|2[H 1 (Ω)]d .
i,j=1
2 i,j=1
∂x j 2
Ω Ω
Considering the closure of C0∞ (Ω) with respect to the norm · H 1 (Ω) we
conclude the assertion for v ∈ [H01 (Ω)]d .
Using suitable equivalent norms in [H 1 (Ω)]d we now conclude the [H01 (Ω)]d –
ellipticity of the bilinear form a(·, ·).
Corollary 4.16. For v ∈ [H01 (Ω)]d we have
a(v, v) ≥ c v2[H 1 (Ω)]d . (4.39)
Proof. For v ∈ [H01 (Ω)]d we can define by

⎧ ⎡ ⎤2 ⎫1/2
⎪
⎨ d ⎪
⎬
v[H 1 (Ω)]d ,Γ := ⎣ vi (x) dsx ⎦ + |v|2 1 (4.40)
[H (Ω)]
⎪ ⎪
d
⎩ i=1 ⎭
Γ
an equivalent norm in [H 1 (Ω)]d (cf. Theorem 2.6). The assertion then follows
from Lemma 4.14 and by using Korn’s first inequality (4.38).
The ellipticity estimate (4.39) remains valid for vector functions v, where
only some components vi (x) are zero for x ∈ ΓD,i ⊂ Γ . Let
! "
[H01 (Ω, ΓD )]d = v ∈ [H 1 (Ω)]d : γ0int vi (x) = 0 for x ∈ ΓD,i , i = 1, . . . , d .
Then we have
a(v, v) ≥ c v2[H 1 (Ω)]d for all v ∈ [H01 (Ω, ΓD )]d . (4.41)
As for the scalar Laplace operator we can extend the bilinear form a(·, ·)
of the system of linear elasticity by some L2 norm to obtain an equivalent
norm in [H 1 (Ω)]d . This is a direct consequence of Korn’s second inequality,
see [53].
Theorem 4.17 (Korn’s Second Inequality). Let Ω ⊂ Rd be a bounded

domain with piecewise smooth boundary Γ = ∂Ω. Then we have
d
[eij (v, x)]2 dx + v2[L2 (Ω)]d ≥ c v2[H 1 (Ω)]d for all v ∈ [H 1 (Ω)]d .
Ω i,j=1
Using Theorem 2.6 we can introduce further equivalent norms in [H 1 (Ω)]d .

dim (R)
Corollary 4.18. Let R = span{v k }k=1 be the space of all rigid body mo-
tions as given in (1.29). Then we can define
⎧ ⎡ ⎤2 ⎫1/2
⎪
⎨dim (R) ⎪
⎬
d
v[H 1 (Ω)]d ,Γ := ⎣ ⎦
v k (x) v(x)dx + 2
[eij (v, x)] dx
⎪
⎩ k=1 ⎪
⎭
Ω i,j=1Ω
as an equivalent norm in [H 1 (Ω)]d .
4.2.1 Dirichlet Boundary Value Problem

The Dirichlet boundary value problem of linear elasticity reads
−µ∆u(x)−(λ+µ)grad div u(x) = f (x) for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ.
As for the scalar potential equation we define the solution manifold
! "
Vg := v ∈ [H 1 (Ω)]d : γ0int vi (x) = gi (x) for x ∈ Γ, i = 1, . . . , d
where V0 = [H01 (Ω)]d . Then we have to find u ∈ Vg such that

a(u, v) = f , vΩ (4.42)
is satisfied for all v ∈ V0 . Since the bilinear form a(·, ·) is bounded (cf. (4.36))
and [H01 (Ω)]d –elliptic (cf. (4.39)) we conclude the unique solvability of the
variational problem (4.42) by applying Theorem 3.8. Moreover, the unique
solution of (4.42) satisfies
u[H 1 (Ω)]d ≤ c1 f [H −1 (Ω)]d + c2 g[H 1/2 (Γ )]d .
For the solution u ∈ [H 1 (Ω)]d of the Dirichlet boundary value problem we now
compute the associated boundary stress γ1int u ∈ [H −1/2 (Γ )]d as the solution
of the variational problem
γ1int u, wΓ = a(u, Ew) − f , EwΩ
for all w ∈ [H 1/2 (Γ )]d . Here, E : H 1/2 (Γ ) → H 1 (Ω) is the extension operator
which is applied to the components wi ∈ H 1/2 (Ω). Note that

E
γ1int u[H −1/2 (Γ )]d ≤ cIT f [H -−1 (Ω)]d + |u| 1 .
1 − 2ν [H (Ω)]
d
Lemma 4.19. Let u ∈ [H 1 (Ω)]d be the weak solution of the Dirichlet bound-
ary value problem
−µ∆u(x) − (λ + µ)grad div u(x) = 0 for x ∈ Ω, u(x) = g(x) for x ∈ Γ.
Then we have
a(u, u) ≥ c γ1int u2[H −1/2 (Γ )]d . (4.43)
Proof. The associated conormal derivative γ1int u ∈ [H −1/2 (Γ )]d is defined as

the unique solution of the variational problem
γ1int u, wΓ = a(u, Ew) for all w ∈ [H 1/2 (Γ )]d .
As in the proof of Lemma 4.13 we have

⎛ ⎞1/2 ⎛ ⎞1/2

2E ⎝
|a(u, Ew)| ≤ e(u, x)22 dx⎠ ⎝ e(Ew, x)22 dx⎠
1 − 2ν
Ω Ω
and
e(Ew, x)22 dx ≤ |Ew|2[H 1 (Ω)]d .
Ω
Moreover,

d
1+ν
e(u, x)22 dx ≤ [eij (u, x)]2 dx ≤ a(u, u).
E
Ω Ω i,j=1
Applying the inverse trace theorem (Theorem 2.22) this gives
γ1int u, wΓ
γ1int u[H −1/2 (Γ )]d = sup
0=w∈[H 1/2 (Γ )]d w[H 1/2 (Γ )]d
a(u, Ew) .
= sup ≤ c a(u, u).
0=w∈[H 1/2 (Γ )]d w[H 1/2 (Γ )]d
Note that the constant c in the estimate (4.43) tends to zero when ν → 12 .

4.2.2 Neumann Boundary Value Problem
For the solvability of the Neumann boundary value problem
−µ∆u(x)−(λ+µ)grad div u(x) = f (x) for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ
we have to assume the solvability conditions (1.31),


v k (x) f (x)dx + γ0int v k (x) g(x)dsx = 0 for all v k ∈ R
Ω Γ
where v k are the rigid body motions (cf. (1.29)). On the other hand, the
solution of the Neumann boundary value problem is only unique up to the
rigid body motions. To fix the rigid body motions, we formulate appropriate
scaling conditions. For this we define
⎧ ⎫
⎨ ⎬
[H∗1 (Ω)]d = v ∈ [H 1 (Ω)]d : v k (x) v(x)dx = 0 for all v k ∈ R .
⎩ ⎭
Ω
The weak formulation of the Neumann boundary value problem is to find

u ∈ [H∗1 (Ω)]d such that
a(u, v) = f , vΩ + g, γ0int vΓ (4.44)
is satisfied for all v ∈ [H∗1 (Ω)]d . Using Corollary 4.18 we can establish the
[H∗1 (Ω)]d –ellipticity of the bilinear form a(·, ·) and therefore we can conclude
the unique solvability of the variational problem (4.44) in [H∗1 (Ω)]d .
By introducing Lagrange multipliers we can formulate the scaling condi-
tions conditions of [H∗1 (Ω)]d as side conditions in a saddle point problem.
Then we have to find u ∈ [H 1 (Ω)]d and λ ∈ Rdim(R) such that
R
dim
a(u, v) + λk v k (x) v(x)dx = f , vΩ + g, γ0int vΓ
k=1 Ω (4.45)

v (x) u(x)dx =0
Ω
is satisfied for all v ∈ [H 1 (Ω)]d and = 1, . . . , dim(R). When choosing as test

functions v ∈ R we then find from the solvability conditions (1.31)
R
dim
λk v k (x) v (x)dx = 0 for = 1, . . . , dim(R).
k=1 Ω
Since the rigid body motions are linear independent, we obtain λ = 0. In-
serting this result into the second equation in (4.45), and eliminating the
Lagrange multiplier λ we finally obtain a modified variational problem to find
u ∈ [H 1 (Ω)]d such that
R
dim
a(u, v) + v k (x) u(x)dx v k (x) v(x)dx = f , vΩ + g, γ0int vΓ
k=1 Ω Ω
(4.46)
is satisfied for all v ∈ [H 1 (Ω)]d .
4.3 Stokes Problem 79
The extended bilinear form of the modified variational formulation (4.46) is

[H 1 (Ω)]d –elliptic (cf. Corollary 4.18). Hence there exists a unique solution
u ∈ [H 1 (Ω)]d of the variational problem (4.46) for any given f ∈ [H −1 (Ω)]d
−1/2
and g ∈ [H d
(Γ )] . If the solvability conditions (1.31) are satisfied, then
we have u ∈ [H∗1 (Ω)]d , i.e. the variational problems (4.46) and (4.44) are
equivalent.
If u ∈ [H∗1 (Ω)]d is a weak solution of the Neumann boundary value prob-
lem, then we can define
R
dim
:= u +
u αk v k ∈ [H 1 (Ω)]d
k=1
which is also a solution of the Neumann boundary value problem.
4.2.3 Mixed Boundary Value Problems
We now consider a boundary value problem with boundary conditions of

mixed type,
−µ∆u(x) − (λ + µ)grad div u(x) = f (x) for x ∈ Ω,

γ0int ui (x) = gD,i (x) for x ∈ ΓD,i ,
(γ1int u)i (x) = gN,i (x) for x ∈ ΓN,i
where Γ = Γ D,i ∪ Γ N,i and meas(ΓD,i ) > 0 for i = 1, . . . , d. The associated

variational formulation is to find u ∈ [H 1 (Ω)]d with γ0int ui (x) = gD,i (x) for
x ∈ ΓD,i such that

d
a(u, v) = f , vΩ + gN,i , γ0int vi ΓN,i (4.47)
i=1
is satisfied for all v ∈ [H01 (Ω, ΓD )]d .

Using (4.41) we conclude the [H01 (Ω, ΓD )]d –ellipticity of the bilinear form
a(·, ·) and therefore the unique solvability of the variational problem (4.47).
4.3 Stokes Problem
Next we consider the Dirichlet boundary value problem for the Stokes system
(1.38),
−µ∆u(x)+∇p(x) = f (x), div u(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ.
Due to (1.39) we have to assume the solvability condition


[n(x)] g(x)dsx = 0. (4.48)
Γ
Note that the pressure p is only unique up to an additive constant. However,

as for the Neumann boundary value problem for the potential equation we
can introduce an appropriate scaling condition to fix this constant. For this
we define ⎧ ⎫
⎨ ⎬
L2,0 (Ω) = q ∈ L2 (Ω) : q(x)dx = 0 .
⎩ ⎭
Ω
To derive a variational formulation for the solution u of the Dirichlet boundary

value problem of the Stokes system, we consider Green’s first formula (1.41),

a(u, v) = p(x)div v(x)dx + f , vΩ + t(u, p), γ0int vΓ .
Ω
Due to div u(x) = 0 for x ∈ Ω, the bilinear form (1.42) is given as

d
a(u, v) := 2µ eij (u, x)eij (v, x)dx.
Ω i,j=1
Hence we have to find u ∈ [H 1 (Ω)]d satisfying u(x) = g(x) for x ∈ Γ and

p ∈ L2,0 (Ω) such that

a(u, v) − p(x) div v(x)dx = f , vΩ ,
Ω
(4.49)
q(x) div u(x) dx =0
Ω
is satisfied for all v ∈ [H01 (Ω)]d and q ∈ L2,0 (Ω).

Let ug ∈ [H 1 (Ω)]d be any arbitrary but fixed extension of the given Dirich-
let datum g ∈ [H 1/2 (Γ )]d . It remains to find u0 ∈ [H01 (Ω)]d and p ∈ L2,0 (Ω)
such that

a(u0 , v) − p(x) div v(x)dx = f , vΩ − a(ug , v),
Ω
(4.50)
q(x) div u0 (x) dx = −div ug , qΩ
Ω
is satisfied for all v ∈ [H01 (Ω)]d and q ∈ L2,0 (Ω).

To investigate the unique solvability of the saddle point problem (4.50)
we have to check the assumptions of Theorem 3.11. The bilinear form a(·, ·) :
[H01 (Ω)]d × [H01 (Ω)]d → R induces an operator A : [H01 (Ω)]d → [H −1 (Ω)]d .
The bilinear form

b(v, q) := q(x) div v(x) dx for v ∈ [H01 (Ω)]d , q ∈ L2 (Ω)
Ω
induces an operator B : [H01 (Ω)]d → L2 (Ω). Note that

|b(v, q)| = q(x) div v(x) dx

Ω
≤ qL2 (Ω) div vL2 (Ω) ≤ qL2 (Ω) v[H 1 (Ω)]d .
We further have

ker B := v ∈ [H01 (Ω)]d : div v = 0 ⊂ [H01 (Ω)]d .
Applying Korn’s first inequality (4.38) this gives
d
a(v, v) = 2µ [eij (v, x)]2 dx ≥ µ |v|2[H 1 (Ω)]d
Ω i,j=1
for all v ∈ [H01 (Ω)]d . Using the equivalent norm (4.40) we then find the
[H01 (Ω)]d –ellipticity of the bilinear form a(·, ·),
a(v, v) ≥ c v2[H 1 (Ω)]d for all v ∈ [H01 (Ω)]d .
Due to ker B ⊂ [H01 (Ω)]d we also have the ker B–ellipticity of the bilinear
form a(·, ·). It remains to prove the stability condition (3.25),

q(x) div v(x)dx
Ω
cS qL2 (Ω) ≤ sup for all q ∈ L2,0 (Ω). (4.51)
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
This is a direct consequence of Theorem 2.17:
Lemma 4.20. Let Ω ⊂ Rd be a bounded and connected Lipschitz domain.
Then there holds the stability condition (4.51) .
Proof. For q ∈ L2,0 (Ω) we have ∇q ∈ [H −1 (Ω)]d satisfying, by using Theorem
2.17,
qL2 (Ω) ≤ c ∇q[H −1 (Ω)]d .
Recalling the norm definition in [H −1 (Ω)]d by duality, this gives
1 w, ∇qΩ
qL2 (Ω) ≤ ∇q[H −1 (Ω)]d = sup
c 1
0=w∈[H0 (Ω)] d w[H 1 (Ω)]d

− q(x) div w(x)dx
Ω
= sup .
0=w∈[H01 (Ω)]d w[H 1 (Ω)]d
Hence, choosing v := −w we finally obtain the stability condition (4.51).

Therefore, all assumptions of Theorem 3.11 are satisfied, and hence, the saddle
point problem (4.50) is unique solvable.
The scaling condition to fix the pressure p ∈ L2,0 (Ω) can now be reformu-
lated as for the Neumann boundary value problem for the potential equation.
By introducing a scalar Lagrange multiplier λ ∈ R we may consider the fol-
lowing extended saddle point problem to find u ∈ [H 1 (Ω)]d with u(x) = g(x)
for x ∈ Γ as well as p ∈ L2 (Ω) and λ ∈ R such that

a(u, v) − p(x) div v(x)dx = f , vΩ ,
Ω
q(x) div u(x) dx + λ q(x)dx = 0, (4.52)
Ω Ω
p(x)dx =0
Ω
is satisfied for all v ∈ [H01 (Ω)]d and q ∈ L2 (Ω). Choosing as test function
q ≡ 1 this gives

λ |Ω| = − div u(x) dx = − n(x) g(x)dsx = 0
Ω Γ
and using the solvability condition (4.48) we get λ = 0. Hence we can write
the third equation in (4.52) as

p(x)dx − λ = 0.
Ω
Eliminating the Lagrange multiplier λ we finally obtain a modified saddle

point problem which is equivalent to the variational formulation (4.50) to find
u0 ∈ [H01 (Ω)]d and p ∈ L2 (Ω) such that

a(u0 , v) − p(x) div v(x)dx = f, vΩ
Ω (4.53)
q(x) div u0 (x) dx + p(x)dx q(x)dx = −div ug , qΩ
Ω Ω Ω
is satisfied for all v ∈ [H01 (Ω)]d and q ∈ L2 (Ω) where
f, vΩ := f , vΩ − a(ug , v) for all v ∈ [H01 (Ω)]d
induces f ∈ [H −1 (Ω)]d .
By
Au, vΩ := a(u, v),

Bu, qL2 (Ω) := b(u, q),
v, B pΩ := Bv, pΩ = b(v, p),

Dp, qL2 (Ω) := p(x)dx q(x)dx
Ω Ω
for all u, v ∈ [H01 (Ω)]d and p, q ∈ L2 (Ω) we can define bounded operators
A : [H01 (Ω)]d → [H −1 (Ω)]d ,

B : [H01 (Ω)]d → L2 (Ω),
B : L2 (Ω) → [H −1 (Ω)]d ,
D : L2 (Ω) → L2 (Ω).
Hence we can write the variational problem (4.53) as an operator equation,

& '& ' & '
A −B u0 f
= .
B D p −Bug
Since A is [H01 (Ω)]d –elliptic we find

2 3
u0 = A−1 f + B p
and inserting this into the second equation we obtain the Schur complement
system 0 1
BA−1 B + D p = −Bug − A−1 f. (4.54)
Lemma 4.21. The operator S := BA−1 B + D : L2 (Ω) → L2 (Ω) is bounded

and L2 (Ω)–elliptic.
Proof. For p ∈ L2 (Ω) we have up = A−1 B p ∈ [H01 (Ω)]d which is defined as

a(up , v) = b(v, p) for all v ∈ [H01 (Ω)]d .
Since A is [H01 (Ω)]d –elliptic, and since B is bounded, we obtain

2
1 up [H 1 (Ω)]d ≤ a(up , up ) = b(up , p) ≤ c2 up [H 1 (Ω)]d pL2 (Ω)
cA B
and therefore
cB
2
up [H 1 (Ω)]d ≤ pL2 (Ω) .
cA
1
Applying the Cauchy–Schwarz inequality this gives
Sp, qL2 (Ω) = [BA−1 B + D]p, qL2 (Ω)

= q(x)div up (x)dx + p(x)dx q(x)dx
Ω Ω Ω
≤ cB
2 qL2 (Ω) up [H 1 (Ω)]d + |Ω| pL2 (Ω) qL2 (Ω)
≤ c pL2 (Ω) qL2 (Ω)
and hence we conclude the boundedness of S : L2 (Ω) → L2 (Ω).

For an arbitrary given p ∈ L2 (Ω) we consider the decomposition

1
p = p0 + p(x)dx
|Ω|
Ω
where p0 ∈ L2,0 (Ω) and

⎡ ⎤2

1 ⎣ p(x) dx⎦ .
p2L2 (Ω) = p0 2L2 (Ω) +
|Ω|2
Ω
For p0 ∈ L2,0 (Ω) we can use the stability condition (4.51), the definition of
up0 = A−1 B p0 ∈ [H01 (Ω)]d and the boundedness of A to obtain
b(v, p0 )
cS p0 L2 (Ω) ≤ sup
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
a(up0 , v)
= sup ≤ cA
2 up0 [H 1 (Ω)]d
0=v∈[H01 (Ω)]d v[H 1 (Ω)]d
and therefore
2 2
cA
2 1 cA
2
p0 2L2 (Ω) ≤ up0 2[H 1 (Ω)]d ≤ a(up0 , up0 ) = c b(up0 , p0 ).
cS cA
1 cS
Inserting the definition of up0 = A−1 B p0 ∈ [H01 (Ω)]d we get
1
BA−1 B p0 , p0 L2 (Ω) = Bup0 , p0 L2 (Ω) ≥ p0 2L2 (Ω) .
c
For v ∈ [H01 (Ω)]d we have

B p, vΩ = p(x)div v(x)dx = −v, ∇pΩ = −v, ∇p0 Ω
Ω
and hence
1
BA−1 B p, pL2 (Ω) = BA−1 B p0 , p0 L2 (Ω) ≥ p0 2L2 (Ω) .
c
From this we obtain
Sp, pL2 (Ω) = BA−1 B p, pL2 (Ω) + Dp, pL2 (Ω)

⎡ ⎤2

= BA−1 B p0 , p0 L2 (Ω) + ⎣ p(x) dx⎦
Ω
⎡ ⎤2

1 1
≥ p0 2L2 (Ω) + ⎣ p(x) dx⎦ ≥ min , |Ω|2 p2L2 (Ω) ,
c c
Ω
i.e., S is L2 (Ω)–elliptic.

Applying Theorem 3.4 (Lax–Milgram lemma) we finally obtain the unique
solvability of the operator equation (4.54).

Finally we consider the interior Dirichlet boundary value problem for the
Helmholtz equation,
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ. (4.55)
The related variational formulation is to find u ∈ H 1 (Ω) with γ0int u(x) = g(x)
for x ∈ Γ such that

∇u(x)∇v(x)dx − k 2 u(x)v(x)dx = 0 (4.56)
Ω Ω
is satisfied for all v ∈ H01 (Ω). The bilinear form

2
a(u, v) = ∇u(x)∇v(x) dx − k u(x)v(x)dx
Ω Ω
can be written as
a(u, v) = a0 (u, v) − c(u, v) (4.57)
where the symmetric and bounded bilinear form

a0 (u, v) = ∇u(x)∇v(x) dx for u, v ∈ H 1 (Ω)
Ω
is H01 (Ω)–elliptic, and the bilinear form

c(u, v) = k 2 u(x)v(x)dx = Cu, vΩ
Ω
induces a compact operator C : H 1 (Ω) → H −1 (Ω). Hence, the bilinear form

a(·, ·) : H01 (Ω) × H01 (Ω) → R induces a coercive and bounded operator A :
H01 (Ω) → H −1 (Ω). Therefore it remains to investigate the injectivity of A,
i.e. we have to consider the solvability of the homogeneous Dirichlet boundary
value problem
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω, γ0int u(x) = 0 for x ∈ Γ. (4.58)
Proposition 4.22. If k 2 = λ is an eigenvalue of the interior Dirichlet eigen-

value problem of the Laplace equation,
−∆u(x) = λ u(x) for x ∈ Ω, γ0int (x) = 0 for x ∈ Γ,
then there exist non–trivial solutions of the homogeneous Dirichlet boundary

value problem (4.58).
If k 2 is not an eigenvalue of the Dirichlet problem of the Laplace operator,
then the operator A : H01 (Ω) → H −1 (Ω) which is induced by the bilinear form
(4.57) is injective.
Hence we conclude, that if k 2 is not an eigenvalue of the Dirichlet eigen-

value problem of the Laplace operator the operator A : H01 (Ω) → H −1 (Ω)
which is induced by the bilinear form (4.57) is coercive and injective, and
therefore the variational problem (4.56) admits a unique solution.
4.5 Exercises
4.1 Derive the variational formulation of the following boundary value prob-
lem with nonlinear Robin boundary conditions
−∆u(x)+u(x) = f (x) for x ∈ Ω, γ1int u(x)+[γ0int u(x)]3 = g(x) for x ∈ Γ
and discuss the unique solvability of the variational problem.

4.2 Consider the bilinear form

d d
∂ ∂ ∂
a(u, v) = aji (x) u(x) v(x)dx + bi (x) u(x)v(x)dx
i,j=1
∂x i ∂x j i=1
∂x i
Ω Ω

+ c(x)u(x)v(x)dx
Ω
4.5 Exercises 87
which is related to a uniform elliptic partial differential operator, and where

we assume c(x) ≥ c0 > 0. Formulate a sufficient condition on the coefficients
bi (x) such that the bilinear form a(·, ·) is H 1 (Ω)–elliptic.
4.3 Consider the Dirichlet boundary value problem
−div[α(x)∇u(x)] = f (x) for x ∈ Ω, γ0int u(x) = 0 for x ∈ Γ
where
ε for x ∈ Ω0 ⊂ Ω,
α(x) =
1 for x ∈ Ω\Ω0 .
Prove the unique solvability of the related variational formulation. Discuss the
dependence of the constants on the parameter ε << 1. Can these constants be
improved when using other norms?
4.4 Formulate a sufficient condition on the wave number k such that the
variational formulation of the interior Neumann boundary value problem of
the Helmholtz equation,
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ,
admits a unique solution.

5
Fundamental Solutions
We now consider the scalar partial differential equation (1.10),
(Lu)(x) = f (x) for x ∈ Ω ⊂ Rd ,
with an elliptic partial differential operator of second order,

d
∂ ∂
(Lu)(x) = − aji (x) u(x) .
i,j=1
∂xj ∂xi
The associated conormal derivative (1.7) is

d
∂
γ1int u(x) = nj (x)aji (x) u(x) for x ∈ Γ
i,j=1
∂xi
and Green’s second formula (1.8) reads for the solution u of the partial dif-
ferential equation (1.10) and for an arbitrary test function v

(Lv)(y)u(y)dy = γ1 u(y)γ0 v(y)dsy − γ1int v(y)γ0int u(y)dsy
int int
Ω Γ Γ

+ f (y)v(y)dy.
Ω
If there exists for any x ∈ Ω a function v(y) := U ∗ (x, y) satisfying

(Ly U ∗ )(x, y)u(y)dy = u(x) (5.1)
Ω
then the solution of the partial differential equation (1.10) is given by the
representation formula for x ∈ Ω
90 5 Fundamental Solutions

u(x) = U ∗ (x, y)γ1int u(y)dsy − int U ∗ (x, y)γ int u(y)ds
γ1,y 0 y (5.2)
Γ Γ

+ U ∗ (x, y)f (y)dy.
Ω
Hence we can describe any solution of the partial differential equation (1.10)
just by knowing the Cauchy data [γ0int u(x), γ1int u(x)] for x ∈ Γ .
Due to
u(x) = δ0 (y − x)u(y)dy for x ∈ Ω
Ω
we have to solve a partial differential equation in the distributional sense to

find the solution of (5.1),
(Ly U ∗ )(x, y) = δ0 (y − x) for x, y ∈ Rd . (5.3)
Any solution U ∗ (x, y) of (5.3) is called a fundamental solution.

The existence of a fundamental solution U ∗ (x, y) is essential to derive the
representation formula (5.2), and therefore to formulate appropriate boundary
integral equations to find the complete Cauchy data. For general results on
the existence of fundamental solutions for partial differential operators we
refer to [79, 90]. In particular for partial differential operators with piecewise
constant coefficients the existence of a fundamental solution is ensured. But
here we will only consider the explicit computation of fundamental solutions
for the Laplace operator, for the system of linear elastostatics, for the Stokes
system, and for the Helmholtz operator.
5.1 Laplace Operator

Let us first consider the Laplace operator
(Lu)(x) := −∆u(x) for x ∈ Rd , d = 2, 3.
The corresponding fundamental solution U ∗ (x, y) is the distributional solution

of the partial differential equation
−∆y U ∗ (x, y) = δ0 (y − x) for x, y ∈ Rd .
Since the Laplace operator is invariant with respect to translations and rota-
tions, we can find the fundamental solution as U ∗ (x, y) = v(z) with z := y −x.
Hence we have to solve
−∆v(z) = δ0 (z) for z ∈ Rd . (5.4)
Applying the Fourier transformation (2.14) this gives, when considering the
derivation rule (2.17),
5.1 Laplace Operator 91
1
|ξ|2 v%(ξ) =
(2π)d/2
and therefore
1 1
v%(ξ) = ∈ S (Rd ).
(2π)d/2 |ξ|2
For the Fourier transform v% of a tempered distribution v ∈ S (Rd ) we have
by definition
%
v , ϕL2 (Rd ) = v, ϕ
% L2 (Rd ) for all ϕ ∈ S(Rd ).
Using
ϕ(ξ) = (2π)−d/2 %
eiz,ξ ϕ(z)dz
Rd
it follows that

1 1
%
v , ϕL2 (Rd ) = %
eiz,ξ ϕ(z)dzdξ.
(2π)d |ξ|2
Rd Rd
Since the integral

1
dξ
|ξ|2
Rd
does not exist we can not exchange the order of integration. However, using
∆z eiz,ξ = −|ξ|2 eiz,ξ
we can consider a splitting of the exterior integral, apply integration by parts

and exchange the order of integration, and repeat integration by parts to
obtain

1 1
%
v , ϕL2 (Rd ) = %
eiz,ξ ϕ(z)dzdξ
(2π)d |ξ|2
Rd Rd

1 1 1 1 eiz,ξ
= %
eiz,ξ ϕ(z)dzdξ + −∆z %
ϕ(z)dzdξ
(2π)d |ξ|2 (2π)d |ξ|2 |ξ|2
|ξ|≤1 Rd |ξ|>1 Rd

1 1 1 1
= %
eiz,ξ ϕ(z)dzdξ + %
eiz,ξ [−∆z ϕ(z)] dzdξ
(2π)d |ξ|2 (2π)d |ξ|4
|ξ|≤1 Rd |ξ|>1 Rd
iz,ξ
1 e 1 eiz,ξ
= %
ϕ(z) dξdz + %
[−∆z ϕ(z)] dξdz
(2π)d |ξ|2 (2π)d |ξ|4
Rd |ξ|≤1 Rd |ξ|>1
⎡ ⎤

1 ⎢ eiz,ξ eiz,ξ ⎥
= %
ϕ(z) ⎣ dξ − ∆z dξ ⎦ dz.
(2π)d |ξ|2 |ξ|4
Rd |ξ|≤1 |ξ|>1
In the three–dimensional case d = 3 we use spherical coordinates
ξ1 = r cos ϕ sin θ, ξ2 = r sin ϕ sin θ, ξ3 = r cos θ
for r ∈ (0, ∞), ϕ ∈ (0, 2π), θ ∈ (0, π) to obtain, by using Lemma 2.13,
⎡ ⎤
iz,ξ iz,ξ
1 ⎢ e e ⎥
v(z) = v(|z|) = ⎣ dξ − ∆z dξ ⎦
(2π)3 |ξ|2 |ξ|4
|ξ|≤1 |ξ|>1
⎡ 2π π 1

1 ⎣
= ei|z|r cos θ sin θ dr dθ dϕ
(2π)3
0 0 0
⎤
2ππ ∞
ei|z|r cos θ sin θ
−∆z dr dθ dϕ⎦
r2
0 0 1
⎡ π 1 ⎤
π ∞ i|z|r cos θ
1 ⎣ e sin θ
= ei|z|r cos θ sin θ dr dθ − ∆z dr dθ ⎦ .
(2π)2 r2
0 0 0 1
The transformation u := cos θ gives
π 1
i|z|r cos θ 1 2 i|z|r 3 2
e sin θdθ = ei|z|ru du = e − e−i|z|r = sin |z|r
i|z|r |z|r
0 −1
and therefore
⎡ ⎤
1 ∞
1 ⎣ sin |z|r sin |z|r ⎦
v(z) = dr − ∆z dr .
2π 2 |z|r |z|r3
0 1
Using the transformation s := |z|r we obtain for the first integral
1 |z|
sin |z|r 1 sin s Si(|z|)
I1 := dr = ds = .
|z|r |z| s |z|
0 0
With

sin ax 1 sin ax a cos ax
dx = − + dx
x3 2 x2 2 x2

1 sin ax a cos ax a2 sin ax
=− − − dx
2 x2 2 x 2 x
the computation of the second integral gives
∞ ∞ ∞
sin |z|r 1 sin |z|r 1 cos |z|r |z| sin |z|r
I2 := 3
dr = − 2
− − dr
|z|r 2 |z|r 2 r 1 2 r
1 1
⎡∞ ⎤
1
1 sin |z| 1 |z| ⎣ sin |z|r sin |z|r ⎦
= + cos |z| − dr − dr
2 |z| 2 2 r r
0 0
1 sin |z| 1 |z| 2 π 3
= + cos |z| − − Si(|z|) .
2 |z| 2 2 2
Inserting this and applying the differentiation we obtain

1 Si(|z|) 1 sin |z| 1 π 1
v(z) = − ∆ z + cos |z| − |z| + |z|Si(|z|)
2π 2 |z| 2 |z| 2 4 2

1 1 Si(|z|) 1 sin |z| 1 1
= ∆z |z| + 2 − ∆z + cos |z| + |z|Si(|z|)
8π 2π |z| 2 |z| 2 2
4 56 7
1 1 =0
= .
4π |z|
Hence the fundamental solution of the Laplace operator in three space dimen-
sions is
1 1
U ∗ (x, y) = for x, y ∈ R3 .
4π |x − y|
For the two–dimensional case d = 2 the inverse Fourier transform of the
fundamental solution has to be regularized in some appropriate way [154]. By

1 ϕ(x) − ϕ(0) ϕ(x)
P 2 , ϕL2 (Rd ) = dx + dx
|x| |x|2 |x|2
x∈R2 :|x|≤1 x∈R2 :|x|>1
1
we first define the tempered distribution P ∈ S (Rd ). Then,
|x|2

1 ϕ(ξ) − ϕ(0) ϕ(ξ)
2π v, ϕ
% L2 (R2 ) = P , ϕL2 (R2 ) = dξ + dξ
|ξ|2 |ξ|2 |ξ|2
ξ∈R2 :|ξ|≤1 ξ∈R2 :|ξ|>1
for all ϕ ∈ S(R2 ). With

1 1
ϕ(ξ) = e iz,ξ
%
ϕ(z)dz, ϕ(0) = %
ϕ(z)dz
2π 2π
R2 R2
we then obtain

1 1
(2π)2 v, ϕ
% L2 (R2 ) = [eiz,ξ − 1]ϕ(z)dzdξ
% + %
eiz,ξ ϕ(z)dzdξ.
|ξ|2 |ξ|2
ξ∈R2 :|ξ|≤1 R2 ξ∈R2 :|ξ|>1 R2
Again we can not exchange the order of integration in the second term. How-
ever, as in three–dimensional case we can write
⎡ ⎤

⎢ eiz,ξ − 1 eiz,ξ ⎥
(2π)2 v, ϕ
% L2 (R2 ) = ϕ(z)
% ⎣ 2
dξ + dξ ⎦ dz.
|ξ| |ξ|2
R2 ξ∈R2 :|ξ|≤1 ξ∈R2 :|ξ|>1
With Lemma 2.13 we further have

1 eiz,ξ − 1 eiz,ξ
v(z) = v(|z|) = dξ + dξ
(2π)2 |ξ|2 |ξ|2
ξ∈R2 :|ξ|≤1 ξ∈R2 :|ξ|>1
and using polar coordinates we obtain

1 2π 2 3 ∞ 2π
1 1 ir|z| cos ϕ 1 1 ir|z| cos ϕ
v(z) = e − 1 dϕdr + e dϕdr
(2π)2 r (2π)2 r
0 0 1 0
1 ∞
1 1 1 1
= [J0 (r|z|) − 1]dr + J0 (r|z|)dr
2π r 2π r
0 1
with the first order Bessel function [63, Subsection 8.411],

2π
1
J0 (s) = eis cos ϕ dϕ.
2π
0
Substituting r := s/ we compute

∞
1 J0 (s) − 1 1 J0 (s)
v(z) = ds + ds
2π s 2π s
0
1 ∞ 1
1 J0 (s) − 1 1 J0 (s) 1 1
= ds + ds + ds
2π s 2π s 2π s
0 1
1 C0
=− log |z| − (5.5)
2π 2π
with the constant
1 ∞
1 − J0 (s) J0 (s)
C0 := ds − ds.
s s
0 1
Since any constant satisfies the homogeneous Laplace equation we can ne-
glect constant terms in the definition of the fundamental solution. Hence the
fundamental solution of the Laplace operator in two space dimensions is
1
U ∗ (x, y) = − log |x − y| for x, y ∈ R2 .
2π
In what follows we will describe an alternative approach to compute the fun-
damental solution for the Laplace operator in two space dimensions. From
Lemma 2.13 we know that the solution v(z) depends only on the absolute
value := |z|. For z = 0 the partial differential equation (5.4) can be rewrit-
ten in polar coordinates as
2
∂ 1 ∂
+ v() = 0 for > 0.
∂2 ∂
The general solution of this ordinary differential equation is given by
v() = a log + b, a, b ∈ R.
In particular for b = 0 we have
U ∗ (x, y) = a log |x − y| .
For x ∈ Ω and for a sufficient small ε > 0 let Bε (x) ⊂ Ω be a ball with center
x and with radius ε. For y ∈ Ω\Bε (x) the fundamental solution U ∗ (x, y) is
a solution of the homogeneous Laplace equation −∆y U ∗ (x, y) = 0. Applying
Green’s second formula (1.8) with respect to the bounded domain Ω\Bε (x)
we obtain

∂ ∂
0 = U ∗ (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy + U ∗ (x, y)f (y)dy
∂ny ∂ny
Γ Γ Ω\B ε (x)

∂ ∂
+ U ∗ (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy .
∂ny ∂ny
∂Bε (x) ∂Bε (x)
Taking the limit ε → 0 we first bound

∗ ∂ ∂
U (x, y) u(y)dsy = |a| | log ε| u(y)dsy
∂ny ∂ny
≤ |a| 2π ε | log ε| uC 1 (Ω) .
1
Using ny = (x − y) for y ∈ ∂Bε (x) we have
ε

∂ ∗ (ny , y − x) a
U (x, y)u(y)dsy = a u(y)dsy = − u(y)dsy .
∂ny |x − y|2 ε
∂Bε (x) ∂Bε (x) ∂Bε (x)
The Taylor expansion

u(y) = u(x) + (y − x)∇u(ξ)
with a suitable ξ = x + t(y − x), t ∈ (0, 1) yields

∂ a
U ∗ (x, y)u(y)dsy = −a 2π u(x) − (y − x)∇u(ξ)dy
∂ny ε
where

a

(y − x)∇u(ξ)dy ≤ |a| 2π ε uC 1 (Ω) .
ε
∂Bε (x)
Taking the limit ε → 0 gives

∗ ∂ ∂
−a 2π u(x) = U (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy
∂ny ∂ny
Γ Γ

+ U ∗ (x, y)f (y)dy
Ω
and therefore the representation formula when choosing a = −1/2π.

To summarize, the fundamental solution of the Laplace operator is given
by ⎧
⎪ 1
⎪−
⎨ log |x − y| for d = 2,
2π
U ∗ (x, y) = (5.6)
⎪
⎪ 1 1
⎩ for d = 3.
4π |x − y|
Any solution of the partial differential equation
−∆u(x) = f (x) for x ∈ Ω ⊂ Rd
is therefore given by the representation formula for x ∈ Ω

∂ ∂
u(x) = U ∗ (x, y) u(y)dsy − U ∗ (x, y)u(y)dsy (5.7)
∂ny ∂ny
Γ Γ

+ U ∗ (x, y)f (y)dy.
Ω
Let us now consider the system of linear elastostatics (1.25),
−µ∆u(x) − (λ + µ)grad div u(x) = f (x) for x ∈ Ω ⊂ Rd ,

and the associated second Betti formula (1.30),

d
∂
σij (v, y)ui (y)dy = γ0int v(y) γ1int u(y)dsy (5.8)
i,j=1
∂yj
Ω
Γ
− γ0 u(y) γ1 v(y)dsy + f (y) v(y)dy.
int int
Γ Ω
To derive a representation formula for the components uk (x), x ∈ Ω, we

therefore have to find solutions v k (x, y) satisfying
d
∂
σij (v k (x, y), y)ui (y)dy = uk (x) for x ∈ Ω, k = 1, . . . , d.
i,j=1
∂y j
Ω
Let ek ∈ Rd be the unit vector with ek = δk for k, = 1, . . . , d. Using the

transformation z := y − x we have to solve the partial differential equations,
k = 1, . . . , d,
−µ∆z v k (z) − (λ + µ)gradz divz v k (z) = δ0 (z) ek for z ∈ Rd .
Using (1.32),
λ+µ
v k (z) := ∆[ψ(z)ek ] − grad div [ψ(z)ek ],
λ + 2µ
we have to find the Airy stress function ψ satisfying the Bi–Laplace equation
−µ∆2 ψ(z) = δ0 (z) for z ∈ Rd
or
−µ∆ϕ(z) = δ0 (z), ∆ψ(z) = ϕ(z) for z ∈ Rd .
From the fundamental solution of the Laplace operator we find
⎧
⎪ 1 1
⎪
⎨ − µ 2π log |z|,
⎪ for d = 2,
ϕ(z) =
⎪
⎪ 1 1 1
⎪
⎩ for d = 3.
µ 4π |z|
For d = 2 we have to solve the remaining Poisson equation when using polar
coordinates,
2
∂ 1 ∂ 1 1
2
+ ψ() = − log for > 0,
∂ ∂ µ 2π
with the general solution

1 1 0 2 1
ψ() = − log − 2 + a log + b for > 0, a, b ∈ R.
µ 8π
In particular for a = b = 0 we have
1 1 0 2 1
ψ() = − log − 2 .
µ 8π
Due to ∆2 |z|2 = 0 we then obtain for = |z|

1 1
ψ(z) = − |z|2 log |z| .
µ 8π
For k = 1 we find for v 1 (z)
λ + µ ∂2 λ + µ ∂2
v11 (z) = ∆ψ(z) − ψ(z), v21 (z) = − ψ(z).
λ + 2µ ∂z12 λ + 2µ ∂z1 ∂z2
∂ zi
Using |z| = we obtain
∂zi |z|
∂ 1 1
ψ(z) = − [2zi log |z| + zi ] (i = 1, 2)
∂zi µ 8π

∂2 1 1 z12
ψ(z) = − 2 log |z| + 2 + 1 (i = 1, 2)
∂zi2 µ 8π |z|2
∂2 1 1 z1 z 2
ψ(z) = −
∂z1 ∂z2 µ 4π |z|2
and therefore
2
1 λ + 3µ 1 λ+µ z1 3
v11 (z) = − log |z| + − ,
4π µ(λ + 2µ) 4π µ(λ + 2µ) |z|2 2
1 λ + µ z1 z 2
v21 (z) = .
4π µ(λ + 2µ) |z|2
For k = 2 the computation is almost the same. Since the constants are so-
lutions of the homogeneous system we can neglect them when defining the
fundamental solution. From v k for k = 1, 2 we then find the Kelvin solution
tensor U ∗ (x, y) = (v 1 , v 2 ) with the components

∗ 1 λ+µ λ + 3µ (yk − xk )(y − x )
Uk (x, y) = − log |x − y| δk +
4π µ(λ + 2µ) λ+µ |x − y|2
for k, = 1, 2. Inserting the Lamé constants (1.24) this gives

∗ 1 1 1+ν (yk − xk )(y − x )
Uk (x, y) = (4ν − 3) log |x − y| δk + .
4π E 1 − ν |x − y|2
This is the fundamental solution of linear elastostatics in two space dimensions

which even exists for incompressible materials with ν = 1/2.
For d = 3 we have to solve the Poisson equation, by using spherical coor-
dinates we obtain

1 ∂ 2 ∂ 1 1 1
ψ() = for > 0,
2 ∂ ∂ µ 4π
with the general solution

1 1 1 a
ψ() = + +b , for > 0, a, b ∈ R.
µ 4π 2
For a = b = 0 we have
1 1
ψ(z) = |z| .
µ 8π
For k = 1 we obtain v 1 (z)
λ + µ ∂2
v11 (z) = ∆ψ(z) − ψ(z),
λ + 2µ ∂z12
λ + µ ∂2
v21 (z) = − ψ(z),
λ + 2µ ∂z1 ∂z2
λ+µ ∂
v31 (z) = − ψ(z)
λ + 2µ ∂z1 ∂z3
and with the derivatives

∂ 1 1 zi ∂2 1 1 1 z2
ψ(z) = , 2 ψ(z) = − i3 ,
∂zi µ 8π |z| ∂zi µ 8π |z| |z|
∂2 1 1 zi zj
ψ(z) = − for i = j
∂zi ∂zj µ 8π |z|3
we then find for the components of the solution v 1

1 λ + 3µ 1 1 λ+µ z12
v11 (z) = + ,
8π µ(λ + 2µ) |z| 8π µ(λ + 2µ) |z|3
1 λ + µ z1 z2
v21 (z) = ,
8π µ(λ + 2µ) |z|3
1 λ + µ z1 z3
v31 (z) = .
8π µ(λ + 2µ) |z|3
For k = 2, 3 the computations are almost the same. Hence, the Kelvin solution
tensor is given by U ∗ (x, y) = (v 1 , v 2 , v 3 ) where

∗ 1 λ+µ λ + 3µ δk (yk − xk )(y − x )
Uk (x, y) = +
8π µ(λ + 2µ) λ + µ |x − y| |x − y|3
for k, = 1, . . . , 3. Inserting the Lamé constants (1.24) this gives the funda-
mental solution of linear elastostatics in three space dimensions

∗ 1 1 1+ν δk (yk − xk )(y − x )
Uk (x, y) = (3 − 4ν) + .
8π E 1 − ν |x − y| |x − y|3
Hence we have the fundamental solution of linear elastostatics

∗ 1 1 1+ν (xk − yk )(x − y )
Uk (x, y) = (3 − 4ν)E(x, y)δk +
4(d − 1)π E 1 − ν |x − y|d
(5.9)
for k, = 1, . . . , d with
⎧
⎨ − log |x − y|
⎪ for d = 2,
E(x, y) = 1
⎪
⎩ for d = 3.
|x − y|
Inserting the solution vectors v(y) = U ∗k (x, y) into the second Betti formula
(5.8) this gives the representation formula

uk (x) = U ∗k (x, y) γ1int u(y)dsy − u(y) γ1,yint U ∗ (x, y)ds
k y
Γ Γ

+ f (y) U ∗k (x, y)dy (5.10)
Ω
for x ∈ Ω and k = 1, . . . , d. Thereby, the boundary stress T ∗k (x, y) of the

fundamental solution U ∗k (x, y) is given for almost all y ∈ Γ by applying (1.27)
as
T ∗k (x, y) := γ1,y
int U ∗ (x, y)
k
∂
= λ divy U ∗k (x, y) n(y) + 2µ U ∗ (x, y) + µ n(y) × curly U ∗k (x, y).
∂ny k
Using
1 1 1+ν yk − xk
div U ∗k (x, y) = 2(2ν − 1)
4(d − 1)π E 1 − ν |x − y|d
we then obtain
1 ν yk − xk E ∂
T ∗k (x, y) = − n(y) + U ∗ (x, y)
2(d − 1)π 1 − ν |x − y|d 1 + ν ∂ny k
E
+ n(y) × curly U ∗k (x, y). (5.11)
2(1 + ν)
Obviously, both the fundamental solutions U ∗k (x, y) and the corresponding
boundary stress functions T ∗k (x, y) exist also for incompressible materials with
ν = 1/2.
5.3 Stokes Problem
Next we consider the Stokes system (1.38),
−µ∆u(x) + ∇p(x) = f (x), div u(x) = 0 for x ∈ Ω ⊂ Rd .
For the solution u and for an arbitrary vector field v we obtain from Green’s
first formula (1.41) by using the symmetry a(u, v) = a(v, u) Green’s second
formula
d

∂
−µ∆vi (y) + q(y) ui (y)dy + p(y)div v(y)dy (5.12)
∂yi
Ω i=1 Ω

d
d
= ti (u, p)vi (y)dsy − ti (v, q)ui (y)dsy + f (y) v(y)dy
Γ i=1 Γ i=1 Ω
with the conormal derivative t(u, p) as given in (1.43).

To obtain representation formulae for the components uk (x), x ∈ Ω, for
the velocity field u we have to find solutions v k (x, y) and q k (x, y) such that
d
∂ k
−µ∆vi (x, y) +
k
q (x, y) ui (y)dy = uk (x), divy v k (x, y) = 0
i=1
∂yi
Ω
for x ∈ Ω, k = 1, . . . , d. With the transformation z := y − x we have to solve

for k = 1, . . . , d
−µ∆v k (z) + ∇q k (z) = δ0 (z)ek , div v k (z) = 0 for z ∈ Rd .
The application of the Fourier transform (2.14) gives
1
d
µ |ξ|2 v%jk (ξ) + i ξj q%k (ξ) = δjk (j = 1, . . . , d), i ξj v%jk (ξ) = 0.
(2π)d/2 j=1
In particular for d = 2 and k = 1 we have to solve a linear system,

1
µ |ξ|2 v%11 (ξ) + i ξ1 q%1 (ξ) = ,
2π
µ |ξ|2 v%21 (ξ) + i ξ2 q%1 (ξ) = 0,
iξ1 v%11 (ξ) + iξ2 v%21 (ξ) = 0,
yielding the solution

1 1 1 ξ12 1 1 ξ1 ξ2 i ξ1
v%11 (ξ) = − , v%21 (ξ) = − , q%1 (ξ) = − .
µ 2π |ξ|2 |ξ|4 µ 2π |ξ|4 2π |ξ|2
As for the scalar Laplace equation we obtain

1 1 1 ξ12
v11 (z) = eiz,ξ
− dξ
µ (2π)2 |ξ|2 |ξ|4
R 2
⎡ ⎤
2
1 1 1 1 ∂ ⎣ 1 1
= eiz,ξ 2 dξ + eiz,ξ 4 dξ ⎦
µ (2π)2 |ξ| µ ∂z12 (2π)2 |ξ|
R2 R2
and with (5.5) we have

1 1 1 C0
eiz,ξ dξ = − log |z| − .
(2π)2 |ξ|2 2π 2π
R2
On the other hand,

⎛ ⎞

1 1 ⎠ = − 1 1 1 C0
∆⎝ 2
eiz,ξ
4
dξ 2
eiz,ξ 2 dξ = log |z| +
(2π) |ξ| (2π) |ξ| 2π 2π
R2 R2
implies

1 1 1 0 2 1 C0 2
eiz,ξ dξ = |z| log |z| − |z|2 + |z| + C1 + C2 log |z|
(2π)2 |ξ|4 8π 8π
R2
with some constants C1 , C2 ∈ R. In particular for C1 = C2 = 0 this gives

1 1 C0 1 ∂2 1 2 C0 2
v11 (z) = − log |z| − + |z| log |z| − |z|2
+ |z|
µ 2π 2π µ ∂z12 8π 8π
2

1 1 z 2C0 + 1
= − log |z| + 12 − .
µ 4π |z| 2
Analogous computations yield
⎡ ⎤
2
1 1 ξ1 ξ2 1 ∂ ⎣ 1 1
v21 (z) = − eiz,ξ 4 dξ = eiz,ξ 4 dξ ⎦
µ (2π)2 |ξ| µ ∂z1 ∂z2 (2π)2 |ξ|
R2 R2

1 ∂2 1 2 C0 2 1 1 z1 z2
= |z| log |z| − |z|2 + |z| =
µ ∂z1 ∂z2 8π 8π µ 4π |z|2
and
⎡ ⎤

i ξ1 ∂ ⎣ 1 1
q 1 (z) = − eiz,ξ dξ = − eiz,ξ dξ ⎦
(2π)2 |ξ|2 ∂z1 (2π)2 |ξ|2
R2 R2

∂ 1 C0 1 ∂
=− − log |z| − = log |z|.
∂z1 2π 2π 2π ∂z1
For d = 2 and k = 2 the computations are almost the same. Neglecting the
constants we finally have the fundamental solution for the Stokes system in
two space dimensions,

∗ 1 1 (yk − xk )(y − x )
Uk (x, y) = − log |x − y| δk + (5.13)
4π µ |x − y|2
1 yk − xk
Q∗k (x, y) = (5.14)
2π |x − y|2
and k, = 1, 2.
For d = 3 we obtain in the same way the fundamental solution for the
Stokes system as

∗ 1 1 δk (yk − xk )(y − x )
Uk (x, y) = +
8π µ |x − y| |x − y|3
1 yk − xk
Q∗k (x, y) =
4π |x − y|3
and k, = 1, . . . , 3.
Comparing the above results with the fundamental solution of the system
of linear elasticity we obtain the equality for
1 1+ν 1
= , (3 − 4ν) = 1
E 1−ν µ
and therefore for
1
ν =
, E = 3µ .
2
The fundamental solution of the linear elasticity system with incompressible
material therefore coincides with the fundamental solution of the Stokes sys-
tem.
Inserting the fundamental solutions v(y) = U ∗k (x, y) and q(y) = Q∗k (x, y)
into the second Greens formula (5.12) this gives the representation formulae

uk (x) = U k (x, y) t(u(y), p(y))dsy − u(y) t(U ∗k (x, y), Q∗k (x, y))dsy
∗
Γ Γ

+ f (y) U ∗k (x, y)dy (5.15)
Ω
for x ∈ Ω and k = 1, . . . , d. Hereby the conormal derivative T ∗k (x, y) is defined

via (1.43) for almost all y ∈ Γ by
T ∗k (x, y) = t(U ∗k (x, y), Q∗k (x, y))
∂
= −Q∗k (x, y)n(x) + 2µ U ∗ (x, y) + µ n(x) × curl U ∗k (x, y)
∂ny k
1 yk − xk ∂
= − n(x) + 2µ U ∗ (x, y) + µ n(x) × curl U ∗k (x, y).
2(d − 1)π |x − y|d ∂ny k
Hence the boundary stress (1.43) of the fundamental solution of the Stokes
system also coincides with the boundary stress (1.27) of the fundamental
solution of the linear elasticity system when choosing ν = 12 and E = 3µ.
It remains to find some appropriate representation formulae for the pres-
sure p. Let us first consider the case d = 2 and the second Green formula
(5.12) where we have to find solutions v 3 (z) and q 3 (z) with z := y − x such
that
−µ∆v 3 (z) + ∇q 3 (z) = 0, div v 3 (z) = δ0 (z) for z ∈ R2 .
By applying the Fourier transformation we obtain the linear system
µ |ξ|2 v%13 (ξ) + i ξ1 q%3 (ξ) = 0,

µ |ξ|2 v%23 (ξ) + i ξ2 q%3 (ξ) = 0,
1
iξ1 v%13 (ξ) + iξ2 v%23 (ξ) =
2π
with the solution
i ξ1 i ξ2 µ
v%13 (ξ) = − , v%23 (ξ) = − , q%3 (ξ) = .
2π |ξ|2 2π |ξ|2 2π
As before we obtain
1 ∂
vi3 (z) = log |z| (i = 1, 2), q 3 (z) = µδ0 (z).
2π ∂zi
Using z := y − x we conclude for x ∈ Ω a representation formula for the
pressure

2
2
p(x) = ti (u, p)vi3 (x, y)dsy − ti (v 3 (x, y), q 3 (x, y))ui (y)dsy
Γ i=1 Γ i=1

2
+ v 3i (x, y)fi (y)dy
Ω i=1
where the conormal derivative (1.43) implies
ti (v 3 (x, y), q 3 (x, y)) = −[µδ0 (y − x) + div v 3 (x, y)]ni (x)

2

+2µ eij (v 3 (x, y), y)nj (y)
j=1
for i = 1, 2, x ∈ Ω and y ∈ Γ . Since v 3 is divergence–free,
2 2
∂ 3 1 ∂2
div v 3 (x, y) = vi (x, y) = log |x − y| = 0,
i=1
∂yi 2π i=1 ∂yi2
we obtain for Γ y = x ∈ Ω
2

ti (v 3 (x, y), q 3 (x, y)) = 2µ eij (v 3 (x, y), y)nj (y).
j=1
Moreover,

3 1 ∂ 3 ∂ 3
eij (v (x, y), y) = v (x, y) + v (x, y)
2 ∂yi j ∂yj i

1 ∂ ∂ ∂ ∂
= log |x − y| + log |x − y|
4π ∂yi ∂yj ∂yj ∂yi
1 ∂ ∂
= log |x − y|
2π ∂yi ∂yj
1 ∂ ∂ ∂ ∗
=− log |x − y| = − Q (x, y).
2π ∂xj ∂yi ∂xj i
Finally we obtain the representation formula for the pressure p, x ∈ Ω,

2
2
∂ ∗
p(x) = ti (u, p)Q∗i (x, y)dy + 2µ Q (x, y)nj (y)ui (y)dsy
i=1 i,j=1
∂xj i
Γ Γ

+ fi (y)Q∗i (x, y)dy . (5.16)
Ω
For d = 3 one may obtain a similar formula, we skip the details.

Finally we consider the Helmholtz
−∆u(x) − k 2 u(x) = 0 for x ∈ Rd , k ∈ R (5.17)
where the computation of the fundamental solution can be done as in the

alternative approach for the Laplace equation.
For d = 3, and by using spherical coordinates we have to solve the partial
differential equation to find v() = v(|x|) = u(x) such that

1 ∂ 2 ∂
− 2 v() − k 2 v() = 0 for > 0.
∂ ∂
With the transformation

1
v() = V ()

this is equivalent to
V () + k 2 V () = 0 for > 0
where the general solution is given by
V () = A1 cos k + A2 sin k
and therefore we obtain

1 cos k sin
v() = V () = A1 + A2 .

When considering the behavior as → 0 we find a fundamental solution of
the Helmholtz equation given by
1 cos k|x − y|
Uk∗ (x, y) = for x, y ∈ R3 .
4π |x − y|
However, it is more common to use a complex combination of the above fun-

damental system to define the fundamental solution by
1 eik|x−y|
Uk∗ (x, y) = for x, y ∈ R3 . (5.18)
4π |x − y|
For d = 2, and by using polar coordinates the Helmholtz equation (5.17) reads
∂2 1 ∂
− v() − v() − k 2 v() = 0 for > 0,
∂2 ∂
or
∂2 ∂
2 v() + v() + k 2 2 v() = 0 for > 0.
∂2 ∂
With the substitution
s 1 ∂
s = k, v() = v( ) = V (s), V (s) = v()
k k ∂
we then obtain a Bessel differential equation of order zero,
s2 V (s) + sV (s) + s2 V (s) = 0 for s > 0. (5.19)
To find a fundamental system of the Bessel differential equation (5.19) we first

consider the ansatz
∞
∞
∞

V1 (s) = vk sk , V1 (s) = vk ksk−1 , V1 (s) = vk k(k − 1)sk−2 .
k=0 k=1 k=2
By inserting this into the differential equation (5.19) we obtain

0 = s2 V1 (s) + sV1 (s) + s2 V1 (s)

∞ ∞ ∞

= vk k(k − 1)sk + vk ksk + vk sk+2
k=2 k=1 k=0
∞
0 1
= vk−2 + k 2 vk sk + v1 s for s > 0
k=2
and thus
1
v1 = 0, vk = − vk−2 for k ≥ 2.
k2
Hence we obtain
1
v2 −1 = 0, v2 = − v2( −1) for = 1, 2, . . .
42
and therefore
(−1)
v2 = v0 for = 1, 2, . . . .
4 (!)2
In particular for v0 = 1 we have
∞
(−1) 2
V1 (s) = 1 + s =: J0 (s)
4 (!)2
=1
which is the first kind Bessel function of order zero. Note that
lim J0 (s) = 1.
s→0
To find a second solution of the fundamental system including a logarithmic

singularity we consider the ansatz
V2 (s) = J0 (s) ln s + W (s) for s > 0.
By using
1
V2 (s) = J0 (s) ln s + J0 (s) + W (s),
s
2 1
V2 (s) = J0 (s) ln s + J0 (s) − 2 J0 (s) + W (s)

s s
we obtain
0 = s2 V2 (s) + sV2 (s) + s2 V2 (s)

0 1
= s2 J0 (s) + sJ0 (s) + s2 J0 (s) ln s
+2sJ0 (s) + s2 W (s) + sW (s) + s2 W (s)
= 2sJ0 (s) + s2 W (s) + sW (s) + s2 W (s)
since J0 (s) is a solution of the Bessel differential equation (5.19). Hence we

have to solve the differential equation
s2 W (s) + sW (s) + s2 W (s) + 2sI0 (s) for s > 0.
With
∞
∞

W (s) = wk sk , J0 (s) = vk ksk−1
k=0 k=1
we have to solve
∞
∞

0 2 1
0= k wk + wk−2 sk + w1 s + 2 vk ksk
k=2 k=1
∞
0 2 1
= k wk + wk−2 + 2kvk sk + [w1 + 2v1 ] s for s > 0.
k=2
Hence we find
w1 = −2v1 = 0,
and
k 2 wk + wk−2 + 2kvk = 0 for k ≥ 2,
i.e.
1
wk = − [wk−2 + 2kvk ] for k ≥ 2.
k2
By using v2 −1 = 0 for ∈ N we then obtain w2 −1 = 0 for ∈ N and
1 0 1 1 1 (−1)
w2 = − w2( −1) + 4v2 = − w2( −1) − .
42 42 4 (!)2
When choosing w0 = 0 we find by induction
(−1) +1 1

w2 = for ∈ N.
4 (!)2 j=1 j
Hence we have
V2 (s) = J0 (s) ln s + W (s)

# $ ⎛ ⎞
∞ ∞

(−1) 2 ⎝ 1
⎠ (−1) s2 .
= 1+ s ln s −
4 (!)2 j=1
j 4 (!)2
=1 =1
Instead of V2 (s) we will use a linear combination of V1 (s) and V2 (s) to define
a second solution of the fundamental system, in particular we introduce the
second kind Bessel function of order zero,
Y0 (s) = [ln 2 − γ] J0 (s) − V2 (s)
where
5.5 Exercises 109
⎡ ⎤
n
1
γ = lim ⎣ − ln n⎦ ≈ 0.57721566490 . . .
n→∞
j=1
j
is the Euler–Mascheroni constant. Note that Y0 (s) behaves like − ln s as s → 0.

The fundamental solution of the Helmholtz equation is then given by
1
Uk∗ (x, y) = Y0 (k |x − y|) for x, y ∈ R2 . (5.20)
2π
5.5 Exercises
5.1 Consider the recursion
1 1 (−1)
w0 = 0, w2 = − w2( −1) − for ∈ N.
42 4 (!)2
Prove by induction that
(−1) +1 1

w2 = for ∈ N.
4 (!)2 j=1 j
5.2 Compute the Green function G(x, y) such that
1
u(x) = G(x, y)f (y)dy for x ∈ (0, 1)
0
is the unique solution of the Dirichlet boundary value problem
−u (x) = f (x) for x ∈ (0, 1), u(0) = u(1) = 0.

6
Boundary Integral Operators
As a model problem we first consider the Poisson equation for d = 2, 3
−∆u(x) = f (x) for x ∈ Ω ⊂ Rd .
The fundamental solution of the Laplace operator is (cf. 5.6)

⎧
⎪ 1
⎨ − 2π log |x − y|
⎪ for d = 2,
∗
U (x, y) =
⎪
⎪ 1 1
⎩ for d = 3,
4π |x − y|
and the solution of the above Poisson equation is given by the representation
formula (5.2)

u(x) = U ∗ (x, y)γ1int u(y)dsy − γ1,y
int U ∗ (x, y)γ int u(y)ds
0 y (6.1)
Γ Γ

∗
+ U (x, y)f (y)dy for x ∈ Ω.
Ω
To derive appropriate boundary integral equations to find the complete

Cauchy data [γ0int u(x), γ1int u(x)] for x ∈ Γ we first have to investigate the
mapping properties of several surface and volume potentials.
6.1 Newton Potential

By
0 f )(x) :=
(N U ∗ (x, y)f (y)dy for x ∈ Rd (6.2)
Ω
we define the volume or Newton potential of a given function f (y), y ∈ Ω.

112 6 Boundary Integral Operators
For ϕ, ψ ∈ S(Rd ) we have

N0 ϕ, ψΩ = 0 ψΩ
ψ(x) U ∗ (x, y)ϕ(y)dydx = ϕ, N
Ω Ω
0 ϕ ∈ S(Rd ). Then we can define the Newton potential

and, therefore, N

N0 : S (R ) → S (Rd ) by
d
N 0 ψΩ
0 f, ψΩ := f, N for all ψ ∈ S(Rd ).
Theorem 6.1. The volume potential N −1 (Ω) → H 1 (Ω) defines a con-

0 : H
tinuous map, i.e.
0 f H 1 (Ω) ≤ c f -−1 .
N (6.3)
H (Ω)
Proof. For ϕ ∈ C0∞ (Ω) we first have

2
|ϕ(ξ)|
%
ϕ2H −1 (Rd ) = dξ
1 + |ξ|2
Rd
% is
where the Fourier transform ϕ

= (2π)− 2 e−ix,ξ ϕ(x)dx.
d
%
ϕ(ξ)
Rd
Due to supp ϕ ⊂ Ω we have
ϕ, vL2 (Rd )

ϕH −1 (Rd ) = sup
0=v∈H 1 (Rd ) vH 1 (Rd )
ϕ, vL2 (Ω)

≤ sup = ϕH
-−1 (Ω) .
0=v∈H 1 (Ω) vH 1 (Ω)
Moreover,

0 ϕ)(x) =
u(x) := (N U ∗ (x, y)ϕ(y)dy for x ∈ Rd .
Ω
Let Ω ⊂ BR (0), and let µ ∈ C0∞ ([0, ∞)) be a non–negative, monotone decreas-
ing cut off function with compact support, and let µ(r) = 1 for r ∈ [0, 2R].
Define

uµ (x) := µ(|x − y|)U ∗ (x, y)ϕ(y)dy for x ∈ Rd .
Ω
Due to µ(|x − y|) = 1 for x, y ∈ Ω we have

6.1 Newton Potential 113
uµ (x) = u(x) for x ∈ Ω
and therefore
uH 1 (Ω) = uµ H 1 (Ω) ≤ uµ H 1 (Rd )
with
uµ 2H 1 (Rd ) = (1 + |ξ|2 ) |%
uµ (ξ)|2 dξ.
Rd
For the computation of the Fourier transform u%µ we obtain

%µ (ξ) = (2π)− 2 e−ix,ξ uµ (x)dx
d
u
Rd

= (2π)− 2 e−ix,ξ µ(|x − y|)U ∗ (x, y)ϕ(y)dydx
d
Rd Rd

−d
= (2π) 2 e−iz+y,ξ µ(|z|)U ∗ (z + y, y)ϕ(y)dydz
Rd Rd

= (2π)− 2 e−iy,ξ ϕ(y)dy e−iz,ξ µ(|z|)U ∗ (z, 0)dz
d
Rd Rd

%
= ϕ(ξ) e−iz,ξ µ(|z|)U ∗ (z, 0)dz .
Rd
Since the function µ(|z|)U ∗ (z, 0) depends only on |z|, we can use Lemma 2.13,
i.e. it is sufficient to evaluate the remaining integral in ξ = (0, 0, |ξ|) .
Let us now consider the case d = 3 only, for d = 2 the further steps are
almost the same. Using spherical coordinates,
z1 = r cos φ sin θ, z2 = r sin φ sin θ, z3 = r cos θ
for r ∈ [0, ∞), φ ∈ [0, 2π), θ ∈ [0, π), we obtain for the remaining integral

1 µ(|z|)
I(|ξ|) = e−iz,ξ dz
4π |z|
Rd
∞ 2ππ
1 µ(r) 2
= e−i|ξ|r cos θ r sin θdθ dφ dr
4π r
0 0 0
∞ π
1
= r µ(r) e−ir|ξ| cos θ sin θ dθ dr.
2
0 0
Using the transformation u = cos θ we get for the inner integral

π 1 1
−ir|ξ| cos θ 1 −ir|ξ|u 2 sin r|ξ|
e sin θ dθ = e−ir|ξ|u du = − e =
ir|ξ| −1 r|ξ|
0 −1
and therefore
∞
1
I(|ξ|) = µ(r) sin r|ξ| dr .
|ξ|
0
For |ξ| > 1 we use the transformation s := r|ξ| to obtain

∞
1 s
I(|ξ|) = µ sin s ds.
|ξ|2 |ξ|
0
Due to 0 ≤ µ(r) ≤ 1 and since µ(r) has compact support, we further conclude
1
|I(|ξ|)| ≤ c1 (R) for |ξ| ≥ 1.
|ξ|2
Note that
(1 + |ξ|2 )2 ≤ 4 |ξ|4 for |ξ| ≥ 1.
Then,

(1 + |ξ|2 )|%
uµ (ξ)|2 dξ = (1 + |ξ|2 )|I(|ξ|)ϕ(ξ)|
% 2
dξ
|ξ|>1 |ξ|>1

1 + |ξ|2 1
≤ [c1 (R)]2 |ϕ(ξ)|
% 2
dξ ≤ 4[c1 (R)]2 |ϕ(ξ)|
% 2
dξ.
|ξ|4 1 + |ξ|2
|ξ|>1 |ξ|>1
For |ξ| ≤ 1 we have

∞
sin r|ξ|
I(|ξ|) = µ(r) dr
|ξ|
0
and therefore
|I(|ξ|)| ≤ c2 (R) for |ξ| ≤ 1.
Hence we have

(1 + |ξ|2 )|%
uµ (ξ)|2 dξ = (1 + |ξ|2 )|I(|ξ|)ϕ(ξ)|
% 2
dξ
|ξ|≤1 |ξ|≤1

1
≤ 2 [c2 (R)]2 |ϕ(ξ)|
% 2
dξ ≤ 4[c2 (R)]2 |ϕ(ξ)|
% 2
dξ.
1 + |ξ|2
|ξ|≤1 |ξ|≤1
Taking the sum this gives


uµ 2H 1 (Rd ) = (1 + |ξ|2 )|%
uµ (ξ)|2 dξ
ξ∈Rd

1 2
≤c |ϕ(ξ)|
% dξ = c ϕ2H −1 (Rd )
1 + |ξ|2
ξ∈Rd
and therefore
0 ϕH 1 (Ω) ≤ c ϕ -−1 .
N H (Ω)
Hence we have
0 f, ϕΩ |
|N 0 ϕΩ |
|f, N
-−1 (Ω) N0 ϕH 1 (Ω)
f H
= ≤ ≤ c f H
-−1 (Ω)
ϕH
-−1 (Ω) ϕH-−1 (Ω) ϕH-−1 (Ω)
for all ϕ ∈ C0∞ (Ω). When taking the closure with respect to the norm
-−1 (Ω) and using a duality arguments gives (6.3).
· H
Theorem 6.2. The volume potential N 0 f is a generalized solution of the par-

tial differential equation

f (x) for x ∈ Ω,

−∆x (N0 f )(x) = f (x) = (6.4)
0 for x ∈ Rd \Ω.
Proof. For ϕ ∈ C0∞ (Rd ) we apply integration by parts, exchange the order of
integration, and using the symmetry of the fundamental solution we obtain

[−∆x (N0 f)(x)]ϕ(x)dx = (N 0 f)(x)[−∆x ϕ(x)]dx
Rd Rd

= U ∗ (x, y)f(y)dy[−∆x ϕ(x)]dx
Rd Rd

= f(y) U ∗ (y, x)[−∆x ϕ(x)]dxdy
Rd Rd

= f(y) [−∆x U ∗ (y, x)]ϕ(x)dxdy
Rd Rd

= f(y) δ0 (x − y)ϕ(x)dxdy
Rd Rd

= f(y)ϕ(y)dy.
Rd
When taking the closure of C0∞ (Rd ) with respect to the norm · H 1 (Rd ) this
shows that the partial differential equation (6.4) is satisfied in the sense of
H −1 (Rd ).

Considering the restriction to the bounded domain Ω ⊂ Rd we further

conclude:
Corollary 6.3. The volume potential N 0 f is a generalized solution of the
partial differential equation
0 f (x) = f (x)
−∆x N for x ∈ Ω.
The application of the interior trace operator
0 f )(x) =
γ0int (N lim 0 f )(
(N x) (6.5)
x→x∈Γ
Ω
defines a linear bounded operator

0 : H
N0 = γ0int N −1 (Ω) → H 1/2 (Γ )
satisfying
N0 f H 1/2 (Γ ) ≤ cN
2 f H
-−1 (Ω)
−1 (Ω).
for all f ∈ H (6.6)
Lemma 6.4. Let f ∈ L∞ (Ω). Then there holds

0 f )(x) =
(N0 f )(x) = γ0int (N U ∗ (x, y)f (y)dy
Ω
for x ∈ Γ as a weakly singular surface integral.

Proof. For an arbitrary given ε > 0 we consider x ∈ Ω and x ∈ Γ satisfying
|x − x| < ε. Then we have

U ∗ (
x, y)f (y)dy − U ∗ (x, y)f (y)dy

Ω y∈Ω:|y−x|>ε

∗ ∗ ∗
≤ x, y) − U (x, y)]f (y)dy +
[U ( U (x, y)f (y)dy ,

y∈Ω:|y−x|>ε y∈Ω:|y−x|≤ε
and

∗ ∗
lim x, y) − U (x, y)]f (y)dy = 0 .
[U (
x→x∈Γ
Ω
y∈Ω:|y−x|>ε
For the remaining term we obtain

∗
x, y)f (y)dy ≤ f L∞ (Ω∩Bε (x))
U ( |U ∗ (
x, y)|dy

y∈Ω:|y−x|≤ε Ω∩Bε (x)

≤ f L∞ (Ω) |U ∗ (
x, y)|dy.
B2ε (
x)
In the case d = 2 we get, by using polar coordinates,

1
|U ∗ (
x, y)|dy = |log |y − x
|| dy
2π
B2ε (
x) |y−
x|<2ε
2π2ε
1
= | log r| r drdϕ = ε2 [1 − 2 log(2ε)] .
2π
0 0
In the same way we find for d = 3, by using spherical coordinates,

1 1
|U ∗ (
x, y)|dy = dy
4π |y − x
|
B2ε (
x) |y−
x|<2ε
2ππ 2ε
1 1 2
= r sin ψ drdψdϕ = 2 ε2 .
4π r
0 0 0
→ x and ε → 0 we finally get the assertion.

Taking the limits x
Lemma 6.5. The operator N1 = γ1int N 0 : H −1 (Ω) → H −1/2 (Γ ) is bounded,

i.e.
0 f H −1/2 (Γ ) ≤ c f -−1
N1 f H −1/2 (Γ ) = γ1int N H (Ω)
−1 (Ω).
is satisfied for all f ∈ H
Proof. First we note that u = N 0 f ∈ H 1 (Ω) is a generalized solution of

the partial differential equation −∆u(x) = f (x) for x ∈ Ω. For an arbitrary
given w ∈ H 1/2 (Γ ) we apply the inverse trace theorem to obtain a bounded
extension Ew ∈ H 1 (Ω) satisfying
EwH 1 (Ω) ≤ cIT wH 1/2 (Γ ) .
Now, using Green’s first formula,

γ1int u, wΓ = ∇u(x)∇Ew(x)dx − f, EwΩ ,
Ω
we get from Theorem 6.1

! "
int
γ1 u, wΓ ≤ uH 1 (Ω) + f H
-−1 (Ω) EwH 1 (Ω)
≤ (c + 1)cIT f H
-−1 (Ω) wH −1/2 (Γ ) .

6.2 Single Layer Potential

Let w ∈ H −1/2 (Γ ) be a given density function. Then we consider the single
layer potential

u(x) := (V w)(x) := U ∗ (x, y)w(y)dsy for x ∈ Ω ∪ Ω c . (6.7)
Γ
Lemma 6.6. The function u(x) = (V w)(x), x ∈ Ω ∪ Ω c , as defined in (6.7)

is a solution of the homogeneous partial differential equation
−∆u(x) = 0 for x ∈ Ω ∪ Ω c .
For w ∈ H −1/2 (Γ ) we have u ∈ H 1 (Ω) satisfying
uH 1 (Ω) = V wH 1 (Ω) ≤ c wH −1/2 (Γ ) .
Proof. For x ∈ Ω ∪ Ω c and y ∈ Γ we notice that the fundamental solution

U ∗ (x, y) is C ∞ . Hence we can exchange differentiation and integration to
obtain

−∆x u(x) = −∆x U ∗ (x, y)f (y)dy = [−∆x U ∗ (x, y)]f (y)dy = 0.
Ω Ω
Moreover, for ϕ ∈ C ∞ (Ω) we have

u(x)ϕ(x)dx = U ∗ (x, y)w(y)dsy ϕ(x)dx
Ω Ω Γ

∗
= w(y) U (x, y)ϕ(x)dx dsy = w(y)(N0 ϕ)(y)dsy
Γ Ω Γ
where
(N0 ϕ)(y) = γ0int U ∗ (x, y)ϕ(x)dx for y ∈ Γ.
Ω
By applying the estimate (6.6) we then obtain

u(x)ϕ(x)dx = w(y)(N0 ϕ)(y)dsy
Ω Γ
≤ wH −1/2 (Γ ) N0 ϕH 1/2 (Γ )
≤ cN
2 wH −1/2 (Γ ) ϕH
-−1 (Ω) .
Taking the closure of C ∞ (Ω) with respect to the norm · H

-−1 (Ω) and using
a duality argument finishes the proof.
6.2 Single Layer Potential 119
The single layer potential (6.7) defines a bounded linear map
V : H −1/2 (Γ ) → H 1 (Ω).
Hence, the application of the interior trace operator to V w ∈ H 1 (Ω) is well

defined. This defines a bounded linear operator
V = γ0int V : H −1/2 (Γ ) → H 1/2 (Γ )
satisfying
V wH 1/2 (Γ ) ≤ cV2 wH −1/2 (Γ ) for all w ∈ H −1/2 (Γ ). (6.8)
Lemma 6.7. Let w ∈ L∞ (Γ ) be given. Then we have the representation

(V w)(x) = γ0int (V w)(x) = U ∗ (x, y)w(y)dsy
Γ
for x ∈ Γ as a weakly singular surface integral.
Proof. For an arbitrary ε > 0 we consider x ∈ Ω and x ∈ Γ satisfying

|x − x
| < ε. Then we have

∗ ∗
U ( x, y)w(y)dsy − U (x, y)w(y)dsy

Γ y∈Γ :|y−x|>ε

∗ ∗ ∗
≤ [U (x, y) − U (x, y)]w(y)dsy + x, y)w(y)dsy ,
U (

y∈Γ :|y−x|>ε y∈Γ :|y−x|≤ε
and for the first expression we obtain

lim [U ∗ (
x, y) − U ∗ (x, y)]w(y)dsy = 0 .
Ωx→x∈Γ
y∈Γ :|y−x|>ε
For the remaining term we have

∗
x, y)w(y)dsy ≤ wL∞ (Γ ∩Bε (x))
U ( |U ∗ (
x, y)|dsy

y∈Γ :|y−x|≤ε Γ ∩Bε (x)

≤ wL∞ (Γ ) |U ∗ (
x, y)|dsy .
Γ ∩Bε (
x)
→ x and ε → 0,
The assertion now follows as in the proof of Lemma 6.4 for x
we skip the details.

In the same way we obtain for the exterior trace
(V w)(x) = γ0ext (V w)(x) := lim (V w)(

x) for x ∈ Γ.
Ω c
x→x∈Γ
Hence we get the jump relation of the single layer potential as
[γ0 V w] := γ0ext (V w)(x) − γ0int (V w)(x) = 0 for x ∈ Γ. (6.9)
6.3 Adjoint Double Layer Potential
For a given density w ∈ H −1/2 (Γ ) we can define V w ∈ H 1 (Ω) which is a

solution of the homogeneous partial differential equation (cf. Lemma 6.6).
Using Lemma 4.4 we can apply the interior conormal derivative to obtain a
bounded linear operator
γ1int V : H −1/2 (Γ ) → H −1/2 (Γ )
satisfying
γ1int V wH −1/2 (Γ ) ≤ c wH −1/2 (Γ ) for all w ∈ H −1/2 (Γ ).
Lemma 6.8. For w ∈ H −1/2 (Γ ) we have the representation
γ1int (V w)(x) = σ(x)w(x) + (K w)(x) for x ∈ Γ
in the sense of H −1/2 (Γ ), i.e.
γ1int V w, vΓ = σw + K w, vΓ for all v ∈ H 1/2 (Γ ).
Here we used the adjoint double layer potential

(K w)(x) := lim int U ∗ (x, y)w(y)ds
γ1,x y (6.10)
ε→0
y∈Γ :|y−x|≥ε
and
1 1
σ(x) := lim dsy for x ∈ Γ. (6.11)
ε→0 2(d − 1)π εd−1
y∈Ω:|y−x|=ε
Proof. For w ∈ H −1/2 (Γ ) the single layer potential V w ∈ H 1 (Ω) is a solution

of the homogeneous partial differential equation. Hence, from Green’s first
formula we find for ϕ ∈ C ∞ (Ω)
6.3 Adjoint Double Layer Potential 121

γ1int u(x)γ0int ϕ(x)dsx = ∇x u(x)∇x ϕ(x)dx
Γ Ω

= ∇x U ∗ (x, y)w(y)dsy ∇x ϕ(x)dx.
Ω Γ
Inserting the definition as weakly singular surface integrals and interchanging

the order of integration this gives

γ1int u(x)γ0int ϕ(x)dsx
Γ ⎛ ⎞

⎜ ⎟
= ∇x ⎝ lim U ∗ (x, y)w(y)dsy ⎠ ∇x ϕ(x)dx
ε→0
Ω y∈Γ :|y−x|≥ε

= w(y) lim ∇x U ∗ (x, y)∇x ϕ(x)dxdsy .
ε→0
Γ x∈Ω:|x−y|≥ε
Using again Green’s first formula we obtain for y ∈ Γ

∗ int U ∗ (x, y)γ int ϕ(x)ds
∇x U (x, y)∇x ϕ(x)dx = γ1,x 0 x
x∈Ω:|x−y|≥ε x∈Γ :|x−y|≥ε

+ int U ∗ (x, y)ϕ(x)ds .
γ1,x x
x∈Ω:|x−y|=ε
The first summand corresponds to the double layer potential operator K as

defined in (6.10). The second term can be written as

int U ∗ (x, y)ϕ(x)ds =
γ1,x int U ∗ (x, y)[ϕ(x) − ϕ(y)]ds
γ1,x
x x
x∈Ω:|x−y|=ε x∈Ω:|x−y|=ε

+ ϕ(y) int U ∗ (x, y)ds
γ1,x x
x∈Ω:|x−y|=ε
with

int U ∗ (x, y)[ϕ(x) − ϕ(y)]ds
γ1,x x

x∈Ω:|x−y|=ε

≤ max |ϕ(x) − ϕ(y)| int U ∗ (x, y)|ds .
|γ1,x x
x∈Ω:|x−y|=ε
x∈Ω:|x−y|=ε
For d = 2 we have

int U ∗ (x, y)|ds ≤
|γ1,x int U ∗ (x, y)|ds
|γ1,x
x x
x∈Ω:|x−y|=ε x∈R2 :|x−y|=ε

1 1
= dsx = 1
2π |x − y|
x∈R2 :|x−y|=ε
while for d = 3

int U ∗ (x, y)|ds ≤
|γ1,x int U ∗ (x, y)|ds
|γ1,x
x x
x∈Ω:|x−y|=ε x∈R3 :|x−y|=ε

1 1
= dsx = 1.
4π |x − y|2
x∈R3 :|x−y|=ε
Taking the limit ε → 0 this gives

int ∗
lim γ1,x U (x, y)[ϕ(x) − ϕ(y)]dsx = 0 .
ε→0
x∈Ω:|x−y|=ε
y−x
For the remaining integral we find by using nx = for x ∈ Ω, |y−x| = ε,
|y − x|

int U ∗ (x, y)ds = − 1 (nx , x − y)
γ1,x x dsx
2(d − 1)π |x − y|d

1 1 1 1
= dsx = dsx .
2(d − 1)π |x − y|d−1 2(d − 1)π εd−1
Taking into account the definitions (6.10) and (6.11) we finally obtain

γ1int u(x)γ0int ϕ(x)dsx
Γ
⎡ ⎤

⎢ int ∗ ⎥
= w(y) ⎣ lim γ1,x U (x, y)γ0int ϕ(x)dsx + γ0int ϕ(y)σ(y)⎦ dsy
ε→0
Γ x∈Γ :|x−y|≥ε

= γ0int ϕ(x) lim int U ∗ (x, y)w(y)ds ds +
γ1,x w(y)σ(y)ϕ(y)dsy
y x
ε→0
Γ y∈Γ :|y−x|≥ε Γ

= [σ(x)w(x) + (K w)(x)]γ0int ϕ(x)dsx .

Γ
6.3 Adjoint Double Layer Potential 123
Let Γ = ∂Ω be at least differentiable within a vicinity of x ∈ Γ . From the

definition (6.11) we then find
1
σ(x) = for almost all x ∈ Γ.
2
The boundary integral operator K which appears in the conormal derivative
of the single layer potential is the adjoint double layer potential. The operator
is linear and bounded, i.e.

K wH −1/2 (Γ ) ≤ cK
2 wH −1/2 (Γ ) for w ∈ H −1/2 (Γ ).
As in the proof of Lemma 6.8 we obtain the following representation of the

exterior conormal derivative of the single layer potential V in the sense of
H −1/2 (Γ ),
γ1ext (V w)(x) = [σ(x) − 1]w(x) + (K w)(x) for x ∈ Γ.
Lemma 6.9. For the conormal derivative of the single layer potential V there
holds the jump relation
[γ1 V w] := γ1ext (V w)(x) − γ1int (V w)(x) = −w(x) for x ∈ Γ (6.12)
in the sense of H −1/2 (Γ ).
Proof. For u = V w and ϕ ∈ C0∞ (Rd ) we first have

[−∆u(x)]ϕ(x)dx = −∆x U ∗ (x, y)w(y)dsy ϕ(x)dx
Rd Rd Γ

= w(y) −∆x U ∗ (x, y)ϕ(x)dxdsy
Γ Rd

= w(y) δ0 (x − y)ϕ(x)dxdsy
Γ Rd

= w(y)γ0int ϕ(y)dsy .
Γ
On the other hand,

[−∆u(x)]ϕ(x)dx = aRd (u, ϕ) = aΩ (u, ϕ) + aΩ c (u, ϕ)
Rd

= γ1int u(x)γ0int ϕ(x)dsx − γ1ext u(x)γ0ext ϕ(x)dsx ,
Γ Γ
and therefore

w(x)ϕ(x)dsx = [γ1int u(x) − γ1ext u(x)]γ0int ϕ(x)dsx
Γ Γ
holds for all ϕ ∈ C0∞ (Rd ). The closure of C0∞ (Rd ) with respect to · H 1/2 (Γ )
and a duality argument then gives the assertion.
6.4 Double Layer Potential

Let v ∈ H 1/2 (Γ ) be a given density function. Then we consider the double
layer potential

u(x) = (W v)(x) := [γ1,y int U ∗ (x, y)]v(y)ds for x ∈ Ω ∪ Ω c . (6.13)
y
Γ
Lemma 6.10. The function u(x) = (W v)(x), x ∈ Ω ∪Ω c , as defined in (6.13)

is a solution of the homogeneous partial differential equation
−∆x u(x) = 0 for x ∈ Ω ∪ Ω c .
For v ∈ H 1/2 (Γ ) we have u ∈ H 1 (Ω) satisfying
uH 1 (Ω) = W vH 1 (Ω) ≤ c vH 1/2 (Γ ) .
Proof. For x ∈ Ω ∪ Ω c and y ∈ Γ we first notice that x = y. Hence we can

interchange differentiation and integration and the first assertion follows from
the properties of the fundamental solution U ∗ (x, y).
For ϕ ∈ C ∞ (Ω) we then have

W v, ϕΩ = int U ∗ (x, y)]v(y)ds ϕ(x)dx
[γ1,y y
Ω Γ

= int
v(y)γ1,y U ∗ (x, y)ϕ(x)dxdsy
Γ Ω

= int (N
v(y)γ1,y 0 ϕΓ .
0 ϕ)(y)dsy = v, γ int N
1
Γ
−1 (Ω) this gives

For f ∈ H
0 f Γ
W v, f Ω = v, γ1int N
and, by applying Corollary 6.3, N0 f ∈ H 1 (Ω) is a solution of the inhomoge-

neous partial differential equation
6.4 Double Layer Potential 125
0 f )(x) = f (x)
−∆x (N for x ∈ Ω.
0 f ∈ H −1/2 (Γ ), and therefore

By Lemma 6.5 we further obtain γ1int N
W v, f Ω
W vH 1 (Ω) = sup
-−1 (Ω)
0=f ∈H
f H
-−1 (Ω)
0 f Γ
v, γ1int N
= sup ≤ c vH 1/2 (Γ ) .

-−1 (Ω) f H
0=f ∈H -−1 (Ω)
The double layer potential (6.13) therefore defines a linear and bounded op-
erator
W : H 1/2 (Γ ) → H 1 (Ω).
When applying the interior trace operator γ0int : H 1 (Ω) → H 1/2 (Γ ) to the
double layer potential u = W v ∈ H 1 (Ω) this declares, for v ∈ H 1/2 (Γ ), a
linear and bounded operator
γ0int W : H 1/2 (Γ ) → H 1/2 (Γ )
satisfying
γ0int W vH 1/2 (Γ ) ≤ c vH 1/2 (Γ ) for v ∈ H 1/2 (Γ ).
Lemma 6.11. For v ∈ H 1/2 (Γ ) we have the representation
γ0int (W v)(x) = [−1 + σ(x)]v(x) + (Kv)(x) for x ∈ Γ (6.14)
where σ(x) is as defined in (6.11) and with the double layer potential

(Kv)(x) := lim int U ∗ (x, y)]v(y)ds
[γ1,y for x ∈ Γ.
y
ε→0
Proof. Let ε > 0 be arbitrary but fixed. For the operator

(Kε v)(x) = int U ∗ (x, y)]v(y)ds
[γ1,y y
we first consider the limit Ω x

→ x ∈ Γ . Hence we assume |
x − x| < ε.
Then,
2 3
x) − (Kε v)(x) =
(W v)( int U ∗ (
γ1,y int U ∗ (x, y) v(y)ds
x, y) − γ1,y y

+ int U ∗ (
[γ1,y x, y)]v(y)dsy
y∈Γ :|y−x|<ε
2 3
= int U ∗ (
γ1,y int U ∗ (x, y) v(y)ds
x, y) − γ1,y y

+ int U ∗ (
[γ1,y x, y)][v(y) − v(x)]dsy
y∈Γ :|y−x|<ε

+ v(x) int U ∗ (
γ1,y x, y)dsy .
y∈Γ :|y−x|<ε
For all ε > 0 we have

2 3
lim int U ∗ (
γ1,y int U ∗ (x, y) v(y)ds = 0
x, y) − γ1,y y
x→x∈Γ
Ω
while the second term can be estimated by

int ∗
x, y)][v(y) − v(x)]dsy
[γ1,y U (

y∈Γ :|y−x|<ε

≤ sup |v(x) − v(y)| int U ∗ (
|γ1,y x, y)|dsy .
y∈Γ :|y−x|<ε
y∈Γ :|y−x|<ε
∈ Ω we further have
For x

int U ∗ (
|γ1,y x, y)|dsy ≤ M .
Γ
Therefore, the second term vanishes when considering the limit ε → 0. For
the computation of the third term we consider
Bε (x) = {y ∈ Ω : |y − x| < ε} .
Then,

γ int U ∗ (
1,y x, y)ds y = int U ∗ (
γ1,y x, y)dsy − int U ∗ (
γ1,y x, y)dsy .
y∈Γ :|y−x|<ε ∂Bε (x) y∈Ω:|y−x|=ε
∈ Bε (x) we
Using the representation formula (6.1) for u = 1 and due to x
obtain
6.4 Double Layer Potential 127

γ1int U ∗ (
x, y)dsy = −1.
∂Bε (x)
Moreover, inserting n(y) = 1ε (y − x) we get

lim lim γ1int U ∗ (
x, y)dsy = lim γ1int U ∗ (x, y)dsy
x→x∈Γ
ε→0 Ω ε→0
y∈Ω:|y−x|=ε y∈Ω:|y−x|=ε

1 (ny , y − x)
= − lim dsy
ε→0 2(d − 1)π |x − y|d
y∈Ω:|y−x|=ε

1 1
= − lim dsy = −σ(x) .

ε→0 2(d − 1)π εd−1
y∈Ω:|y−x|=ε
In the same way we obtain for the exterior trace

γ0ext (W v)(x) = σ(x)v(x) + (Kv)(x) for x ∈ Γ
and therefore the jump relation of the double layer potential,
[γ0 W v] := γ0ext (W v)(x) − γ0int (W v)(x) = v(x) for x ∈ Γ.
Lemma 6.12. For the jump of the conormal derivative of the double layer
potential there holds
[γ1 W v] = γ1ext (W v)(x) − γ1int (W v)(x) = 0 for x ∈ Γ.
Proof. For the double layer potential u(x) = (W v)(x), x ∈ Rn and for ϕ ∈
C0∞ (Rd ) we first have

[−∆u(x)]ϕ(x)dx = −∆x γ1,y int U ∗ (x, y)w(y)ds ϕ(x)dx
y
Rd Rd Γ

= w(y)γ int
1,y −∆x U ∗ (x, y)ϕ(x)dxdsy
Γ Rd

= int
w(y)γ1,y δ0 (x − y)ϕ(x)dxdsy = 0 .
Γ Rd
On the other hand, using Green’s first formula we have

0 = [−∆u(x)]ϕ(x)dx = aRd (u, ϕ) = aΩ (u, ϕ) + aΩ c (u, ϕ)
Rd

= γ1int u(x)γ0int ϕ(x)dsx − γ1ext u(x)γ0ext ϕ(x)dsx
Γ Γ
and taking the closure of C0∞ (Rd )

with respect to the · H 1 (Rd ) norm we
int ext
obtain the assertion from γ0 ϕ = γ0 ϕ.
6.5 Hypersingular Boundary Integral Operator
The conormal derivative of the double layer potential W v for v ∈ H 1/2 (Γ )

defines a bounded operator
γ1int W : H 1/2 (Γ ) → H −1/2 (Γ ).
For
(Dv)(x) := −γ1int (W v)(x) = − lim nx ·∇x (W v)(

x) for x ∈ Γ (6.15)
x→x∈Γ
Ω
we first have
DvH −1/2 (Γ ) ≤ cD
2 vH 1/2 (Γ ) for v ∈ H 1/2 (Γ ). (6.16)
In the two–dimensional case d = 2 the double layer potential reads

1 x − y, ny )
(
(W v)(
x) = lim ∈ Ω.
v(y)dsy for x
2π ε→0 x − y|2
|
For a fixed ε > 0 we can interchange taking the limit x → x ∈ Γ and

computing the conormal derivative to obtain (d = 2)

1 (nx , ny ) (x − y, nx )(x − y, ny )
(Dε v)(x) = − + 2 v(y)dsy .
2π |x − y|2 |x − y|4
In the same way we find for d = 3

1 (nx , ny ) (y − x, ny )(y − x, nx )
(Dε v)(x) = − + 3 v(y)dsy .
4π |x − y|3 |x − y|5
However, when taking the limit ε → 0 for x ∈ Γ the integrals does not exist as
Cauchy principal value. As a generalization of the Cauchy integral we there-
fore call D a hypersingular boundary integral operator. To find an explicit
representation of D we therefore have to introduce a suitable regularisation.
Inserting u0 (x) ≡ 1 into the representation formula (6.1) this gives

1 = − γ1,y int U ∗ ( ∈ Ω.
x, y)dsy for x
Γ
Hence we have
∇x (W u0 )( ∈ Ω,
x) = 0 for x
and therefore
(Du0 )(x) = 0 for x ∈ Γ. (6.17)
6.5 Hypersingular Boundary Integral Operator 129
Moreover we can write

(Dv)(x) = − lim nx · ∇x int U ∗ (
γ1,y x, y)[v(y) − v(x)]dsy for x ∈ Γ.
x→x∈Γ
Ω
Γ
If the density v is continuous, we can obtain for the hypersingular boundary

integral operator D the representation

(Dv)(x) = − γ1,x int γ int U ∗ (x, y)[v(y) − v(x)]ds for x ∈ Γ
1,y y
Γ
as a Cauchy principal value integral.

In what follows we will describe alternative representations of the bilinear
form which is induced by the hypersingular boundary integral operator D,

Du, vΓ = − v(x)γ1,x int int U ∗ (x, y)u(y)ds ds .
γ1,y y x
Γ Γ
In the two–dimensional case d = 2 we assume that Γ = ∂Ω is piecewise

smooth,
*p
Γ = Γk ,
k=1
where each part Γk is given by a local parametrization

y1 (t)
Γk : y = y(t) = for t ∈ (tk , tk+1 ). (6.18)
y2 (t)
Moreover, /
dsy = [y1 (t)]2 + [y2 (t)]2 dt,
and the exterior normal vector is given by

1 y2 (t)
n(y) = . for y ∈ Γk .
[y1 (t)]2 + [y2 (t)]2 −y1 (t)
For x ∈ R2 , the rotation of a scalar function v is defined as

⎛ ⎞
∂
v(x)
⎜ ∂x2 ⎟
curl v(x) := ⎜
⎝
⎟.
⎠
∂
− v(x)
∂x1
If v(x), x ∈ Γk , is a given function, we may consider an appropriate extension
v into a neighborhood of Γk , in particular we may define
v( = x + (
x) = v(x) for x x − x, n(x))n(x).
Then, for x ∈ Γk we can introduce

∂ ∂
curlΓk v(x) := n(x) · curl v(x) = n1 (x) v(x) − n2 (x) v(x)
∂x2 ∂x1
and we obtain

∂ ∂
curlΓk v(y)dsy = n1 (y) v(y) − n2 (y) v(y) dsy
∂y2 ∂y1
Γk Γk
tk+1
∂ ∂
= y2 (t) v(y(t)) + y1 (t) v(y(t)) dt
∂y2 ∂y1
tk
tk+1
d
= v(y(t)) dt,
dt
tk
i.e. curlΓk v does not depend on the chosen extension v.

Lemma 6.13. Let Γk be an open boundary part which is given by a local
parametrization (6.18) with continuously differentiable functions yi (t), i =
1, 2. If v and w are continuously differentiable, then we have the formula of
integration by parts, i.e.

t
v(y) curlΓk w(y) dsy = − curlΓk v(y) w(y) dsy + v(y(t))w(y(t))|tk+1
k
.
Γk Γk
Proof. The assertion follows from

tk+1
d t=t
curlΓk [v(y)w(y)]dsy = [v(y(t))w(y(t))]dt = [v(y(t))w(y(t))]t=tk+1
dt k
Γk tk
through differentiation by the product rule.

For a function v which is defined on a closed curve Γ we define
curlΓ v(x) := curlΓk v(x) for x ∈ Γk , k = 1, . . . , p.
As a consequence of Lemma 6.13 we then have:
Corollary 6.14. Let Γ be a piecewise smooth closed curve. If v and w are
piecewise continuously differentiable, then

p
t
v(y) curlΓ w(y) dsy = − curlΓ v(y) w(y) dsy + v(y(t))w(y(t))|tk+1
k
.
Γ Γ k=1
If in addition v and w are globally continuous, then

v(y) curlΓ w(y) dsy = − curlΓ v(y) w(y) dsy .
Γ Γ
By applying integration by parts we can rewrite the bilinear form which is

induced by the hypersingular boundary integral operator D as a bilinear form
which is induced by the single layer potential V . In case of the two–dimensional
Laplace operator this relation already goes back to [101].
Theorem 6.15. Let Γ be a piecewise smooth closed curve and let u and v be
globally continuous on Γ . Moreover, let u and v be continuously differentiable
on the parts Γk . Then we can rewrite the bilinear form of the hypersingular
boundary integral operator D as

1
Du, vΓ = − curlΓ v(x) log |x − y| curlΓ u(y)dsy dsx . (6.19)
2π
Γ Γ
Proof. The hypersingular boundary integral operator D is defined as the neg-

ative conormal derivative of the double layer potential W , see (6.15). For
∈ Ω we have
x

1 ∂
w(
x) = (W u)(x) = − u(y) log |
x − y|dsy .
2π ∂ny
Γ
∈ Ω and y ∈ Γ it follows that x

Since x = y. With
∂ yi − x
i i − yi
x ∂
log |
x − y| = 2
=− 2
=− log |
x − y|
∂yi |
x − y| |
x − y| ∂
xi
we obtain

∂ ∂ ∂
log |
x − y| = −n(y) · ∇y log |
x − y| .
∂
xi ∂ny ∂yi
Due to ∆y log |
x − y| = 0 for y = x we further get

∂
curlΓ,y log |
x − y| =
∂y1
∂ ∂ ∂ ∂
= n1 (y) log |
x − y| − n2 (y) log |
x − y|
∂y2 ∂y1 ∂y1 ∂y1
∂ ∂ ∂ ∂
= n1 (y) log |
x − y| + n2 (y) log |
x − y|
∂y2 ∂y1 ∂y2 ∂y2

∂
= n(y) · ∇y log |
x − y| ,
∂y2
and
∂ ∂
curlΓ,y log |
x − y| = −n(y) · ∇y log |
x − y| .
∂y2 ∂y1
Hence we can write the partial derivatives of the double layer potential for a
globally continuous function u by applying integration by parts as

∂ 1 ∂ ∂
x) = −
w( u(y) log |
x − y| dsy
∂
x1 2π ∂
x1 ∂ny
Γ

1 ∂
= u(y)n(y) · ∇y log |
x − y| dsy
2π ∂y1
Γ

1 ∂
=− u(y) curlΓ,y log |x − y| dsy
2π ∂y2
Γ

1 ∂
= curlΓ u(y) log |
x − y| dsy ,
2π ∂y2
Γ
and
∂ 1 ∂
x) = −
w( curlΓ u(y) log |
x − y| dsy .
∂
x2 2π ∂y1
Γ
For the normal derivative of the double layer potential we then obtain
n(x) · ∇x w(

x) =

1 ∂ ∂
= curlΓ u(y) n1 (x) log |
x − y| − n2 (x) log |
x − y| dsy
2π ∂y2 ∂y1
Γ

1 ∂ ∂
= lim curlΓ u(y) n1 (x) log |
x − y| − n2 (x) log |
x − y| dsy
2π ε→0 ∂y2 ∂y1
y∈Γ,|y−x|≥ε
and taking the limit Ω x

→ x ∈ Γ this gives
∂
w(x) =
∂nx

1 ∂ ∂
= lim curlΓ u(y) n1 (x) log |x − y| − n2 (x) log |x − y| dsy
2π ε→0 ∂y2 ∂y1

1 ∂ ∂
= − lim curlΓ u(y) n1 (x) log |x − y|−n2 (x) log |x − y| dsy
2π ε→0 ∂x2 ∂x1

1
= − lim curlΓ u(y) curlΓ,x log |x − y| dsy .
2π ε→0
Therefore,

∂ 1
v(x) w(x)dsx = − v(x) lim curlΓ u(y) curlΓ,x log |x − y| dsy dsx
∂nx 2π ε→0
Γ x∈Γ y∈Γ :|y−x|≥ε

1
=− curlΓ u(y) lim v(x) curlΓ,x log |x − y| dsx dsy
2π ε→0
y∈Γ x∈Γ :|x−y|≥ε

1
= curlΓ u(y) lim curlΓ v(x) log |x − y| dsx dsy ,
2π ε→0
y∈Γ x∈Γ :|x−y|≥ε
from which we finally obtain (6.19).

The representation of the hypersingular boundary integral operator D via
integration by parts can be applied correspondingly to the three–dimensional
case [50]. Let
*p
Γ = Γk
k=1
be a piecewise smooth surface where each piece Γk can be described via a

parametrization
⎛ ⎞
y1 (s, t)
y ∈ Γk : y(s, t) = ⎝ y2 (s, t) ⎠ for (s, t) ∈ τ
y3 (s, t)
where τ is some reference element. The rotation or curl of a vector–valued

function v is defined as
curl v := ∇ × v(x) for x ∈ R3 .
If u is a scalar function given on Γk , then
curlΓk u(x) := n(x) × ∇

u(x) for x ∈ Γk
is a suitable extension of the given u on Γk

defines the surface curl, where u
into a three–dimensional neighborhood of Γk . Finally we introduce
curlΓk v(x) := n(x) · curl v(x) for x ∈ Γk .
Lemma 6.16. Let Γ be a piecewise smooth closed Lipschitz surface in R3 .

Assume that each surface part Γk is smooth having a piecewise smooth bound-
ary curve ∂Γk . Let u and v be globally continuous, but locally bounded and
smooth on each Γk . Then, applying integration by parts,

curlΓ u(x) · v(x) dsx = − u(x)curlΓ v(x)dsx .
Γ Γ
Proof. Using the product rule
∇ × [
u(x)v(x)] = ∇
u(x) × v(x) + u
(x) [∇ × v(x)]
we obtain

curlΓk u(x) · v(x)dsx = [n(x) × ∇
u(x)] · v(x)dsx
Γk Γk

= u(x) × v(x)] · n(x) dsx
[∇
Γk

= [∇ × [
u(x)v(x)] − u
(x) [∇ × v(x)]] · n(x) dsx
Γk

= u(x)v(x)t(x)dσ − u(x) curlΓk v(x)dsx
∂Γk Γk
by applying the integral theorem of Stokes.

Theorem 6.17. Let Γ be a piecewise smooth closed surface, and let u and v
be globally continuous functions defined on Γ which are differentiable on Γk .
Then the bilinear form of the hypersingular boundary integral operator D can
be written as

1 curlΓ u(y) · curlΓ v(x)
Du, vΓ = dsx dsy .
4π |x − y|
Γ Γ
∈Ω
Proof. The proof follows essentially as in the two–dimensional case. For x
and using the definition (6.15) of the hypersingular boundary integral operator
D we have to consider the double layer potential

1 ∂ 1
w(x) := − u(y) dsy .
4π ∂ny |
x − y|
Γ
Using
∂ 1 i − yi
x yi − x
i ∂ 1
= = − = −
∂yi |
x − y| x − y|3
| |
x − y| xi |
∂ x − y|
we obtain for the partial derivatives of the kernel function

∂ ∂ 1 ∂ 1
= −n(y) · ∇y .
xi ∂ny |
∂ x − y| ∂yi |
x − y|
Let ei be the i–th unit vector of R3 . Due to x

= y we can expand the vector
product as

1 1
curly ei × ∇y = ∇y × ei × ∇y
|
x − y| |
x − y|

1 1
= ∇y · ∇y ei − (∇y · ei ) ∇y
|
x − y| |
x − y|

1 ∂ 1 ∂ 1
= ∆y e − ∇y = − ∇y .
|
x − y| i ∂yi |
x − y| ∂yi |
x − y|
When exchanging the order of differentiation and integration we then obtain
the partial derivatives of the double layer potential as

∂ 1 ∂ ∂ 1
x) = −
w( u(y) dsy
∂xi 4π xi ∂ny |
∂ x − y|
Γ

1 ∂ 1
= u(y) ny · ∇y dsy
4π ∂yi |
x − y|
Γ

1 1
=− u(y) ny · curly ei × ∇y dsy
4π |
x − y|
Γ

1 1
=− u(y)curlΓ,y ei × ∇y dsy .
4π |
x − y|
Γ
By using Lemma 6.16 we have

∂ 1 1
w(
x) = curlΓ,y u(y) · ei × ∇y dsy
∂
xi 4π |
x − y|
Γ

1 1
=− ei · curlΓ,y u(y) × ∇y dsy ,
4π |
x − y|
Γ
and hence we can write the gradient of the double layer potential as

1 1
∇x w(
x) = − curlΓ,y u(y) × ∇y dsy
4π |
x − y|
Γ

1 1
= curlΓ,y u(y) × ∇x dsy .
4π |
x − y|
Γ
Multiplying this with the normal vector n(x) this gives

1 1
n(x) · ∇x w(
x) = curlΓ,y u(y) × ∇x · n(x) dsy
4π |
x − y|
Γ

1 1
=− curlΓ,y u(y) · n(x) × ∇x dsy
4π |
x − y|
Γ

1 1
=− lim curlΓ,y u(y) · n(x) × ∇x dsy .
4π ε→0 |
x − y|
Taking the limit Ω x → x ∈ Γ we find

1 1
(Du)(x) = − lim curlΓ,y u(y) · n(x) × ∇x dsy
4π ε→0 |x − y|

1 1
=− lim curlΓ,y u(y) · curlΓ,x dsy .
4π ε→0 |x − y|
For the bilinear form of the hypersingular boundary integral operator we

therefore obtain

1 1
Du, vΓ = − v(x) lim curlΓ,y u(y) · curlΓ,x dsy dsx
4π ε→0 |x − y|
Γ y∈Γ :|y−x|≥ε

1 1
=− lim v(x)curlΓ,y u(y) · curlΓ,x dsx dsy
4π ε→0 |x − y|

1 1
= lim curlΓ,x v(x)curlΓ,y u(y) dsx dsy .
4π ε→0 |x − y|
By using
0 1
curlΓ,x [v(x)curlΓ,y u(y)] = n(x) · ∇x × [v(x)curlΓ,y u(y)]
0 1
= n(x) · ∇x v(x) × curlΓ,y u(y)
= [n(x) × ∇x v(x)] · curlΓ,y u(y)
= curlΓ,x v(x) · curlΓ,y u(y)
we finally conclude the assertion.

6.6 Properties of Boundary Integral Operators

Before proving the ellipticity of the single layer potential V and of the hy-
persingular boundary integral operator D we will derive some basic relations
of boundary integral operators. For this we first consider the representation
∈ Ω,
formula (6.1) for x

x) = U ∗ (
u( int U ∗ (
x, y)γ1int u(y)dsy − γ1,y x, y)γ0int u(y)dsy
Γ Γ

+ U ∗ (
x, y)f (y)dy.
Ω
Taking the limit Ω x

→ x ∈ Γ we find from all properties of boundary and
volume potentials as already considered in this chapter a boundary integral
equation for x ∈ Γ,
6.6 Properties of Boundary Integral Operators 137
γ0int u(x) = (V γ1int u)(x) + [1 − σ(x)]γ0int u(x) − (Kγ0int u)(x) + N0 f (x). (6.20)
The application of the conormal derivative to the function u defined by the

representation formula yields a second boundary integral equation for x ∈ Γ ,
γ1int u(x) = σ(x)γ1int u(x) + (K γ1int u)(x) + (Dγ0int u)(x) + N1 f (x). (6.21)
With (6.20) and (6.21) we have obtained a system of two boundary integral
equations which can be written for x ∈ Γ as
& ' & '& ' & '
γ0int u (1 − σ)I − K V γ0int u N0 f
= + (6.22)
γ1int u D σI + K γ1int u N1 f
where & '

(1 − σ)I − K V
C = (6.23)
D σI + K
is the Calderón projection.
Lemma 6.18. The operator C as defined in (6.23) is a projection, i.e., C = C 2 .
Proof. Let (ψ, ϕ) ∈ H −1/2 (Γ ) × H 1/2 (Γ ) be arbitrary but fixed. The function
x) := (V ψ)(
u( x) − (W ϕ)(
x) ∈Ω
for x
is then a solution of the homogeneous partial differential equation. For the

trace and for the conormal derivative of u we find from the properties of the
boundary potentials for x ∈ Γ
γ0int u(x) = (V ψ)(x) + (1 − σ(x))ϕ(x) − (Kϕ)(x),

γ1int u(x) = σψ(x) + (K ψ)(x) + (Dϕ)(x).
The function u is therefore a solution of the homogeneous partial differential

equation whereas the associated Cauchy data are determined for x ∈ Γ by
[γ0int u(x), γ1int u(x)]. These Cauchy data are therefore solutions of the bound-
ary integral equations (6.20) and (6.21), i.e. for x ∈ Γ
(V γ1int u)(x) = (σI + K)γ0int u(x),

(Dγ0int u)(x) = ((1 − σ)I − K )γ1int u(x).
This is equivalent to
& ' & '& '
γ0int u(x) (1 − σ)I − K V γ0int u(x)
= .
γ1int u(x) D σI + K γ1int u(x)
Inserting
& ' & '& '
γ0int u(x) (1 − σ)I − K V ϕ(x)
=
γ1int u(x) D σI + K ψ(x)
this gives the assertion.

From the projection property C = C 2 we can immediately conclude the
following relations of boundary integral operators.
Corollary 6.19. For all boundary integral operators there hold the relations
V D = (σI + K)((1 − σ)I − K), (6.24)

DV = (σI + K )((1 − σ)I − K ), (6.25)
V K = KV, (6.26)
K D = DK. (6.27)
Note that (6.26) describes the symmetrization of the double layer potential
K, which is in general not self–adjoint, by the single layer potential V . This
property was already described by J. Plemelj in 1911 in the case of the two–
dimensional Laplace operator [112].
From the system (6.22) of boundary integral equations we may also find
a suitable representation of the Newton potential N1 f when assuming the
invertibility of the single layer potential V , see also Subsection 6.6.1.
Lemma 6.20. For the volume potential (N1 f )(x), x ∈ Γ , there holds the
representation
(N1 f )(x) = ([σ − 1]I + K )V −1 (N0 f )(x) .
Proof. Using the first boundary integral equation in (6.22) and assuming the
invertibility of the single layer potential V we first obtain
γ1int u(x) = V −1 (σI + K)γ0int u(x) − V −1 (N0 f )(x) for x ∈ Γ.
Inserting this into the second boundary integral equation of (6.22) we get
γ1int u(x) = (Dγ0int u)(x) + (σI + K )γ1int u(x) + (N1 f )(x)

= (Dγ0int u)(x) + (σI + K )[V −1 (σ(x) + K)γ0int u(x) − V −1 (N0 f )(x)]
+(N1 f )(x)
0 1
= D + (σI + K )V −1 (σI + K) γ0int u(x)
−(σI + K )V −1 (N0 f )(x) + (N1 f )(x)
and therefore the equality
−V −1 (N0 f )(x) = −(σI + K )V −1 (N0 f )(x) + (N1 f )(x) for x ∈ Γ.
From this we immediately find the assertion.

6.6.1 Ellipticity of the Single Layer Potential
By using Theorem 3.4 (Lax–Milgram theorem) we can ensure the invertibility

of the single layer potential V : H −1/2 (Γ ) → H 1/2 (Γ ). Hence we need to
prove the H −1/2 (Γ )–ellipticity of V .
The function
u(x) = (V w)(x) for x ∈ Ω
is a solution of the interior Dirichlet boundary value problem
−∆u(x) = 0 for x ∈ Ω, u(x) = γ0int (V w)(x) = (V w)(x) for x ∈ Γ.
Assuming w ∈ H −1/2 (Γ ) we have u = V w ∈ H 1 (Ω) and by choosing

v ∈ H 1 (Ω) we obtain, by applying Green’s first formula (1.5),

aΩ (u, v) := ∇u(x)∇v(x) dx = γ1int u, γ0int vΓ . (6.28)
Ω
Moreover, inequality (4.17) implies
cint int 2
1 γ1 uH −1/2 (Γ ) ≤ aΩ (u, u) . (6.29)
To obtain a corresponding result for the exterior conormal derivative

γ1ext u ∈ H −1/2 (Γ ) we need to investigate the far field behavior of the single
layer potential (V w)(x) as |x| → ∞. For this we first introduce the subspace
! "
−1/2
H∗ (Γ ) := w ∈ H −1/2 (Γ ) : w, 1Γ = 0 (6.30)
of functions which are orthogonal to the constants.
Lemma 6.21. For y0 ∈ Ω and x ∈ Rd we assume
|x − y0 | > max{1, 2 diam(Ω)}

−1/2
to be satisfied. Let w ∈ H −1/2 (Γ ) for d = 3 and w ∈ H∗ (Γ ) for d = 2,
respectively. For u = V w we then have the bounds
1
|u(x)| = |(V w)(x)| ≤ c1 (w) (6.31)
|x − y0 |
and
1
|∇u(x)| = |∇(V w)(x)| ≤ c2 (w) . (6.32)
|x − y0 |2
Proof. By using the triangle inequality we have for y ∈ Ω
1
|x − y0 | ≤ |x − y| + |y − y0 | ≤ |x − y| + diam(Ω) ≤ |x − y| + |x − y0 |
2
and therefore
1
|x − y0 | .
|x − y| ≥
2
In the case d = 3 we find the estimate (6.31) from

|u(x)| = γ0int U ∗ (x, ·), wΓ
≤ γ0int U ∗ (x, ·)H 1/2 (Γ ) wH −1/2 (Γ ) ≤ cT ||U ∗ (x, ·)||H 1 (Ω) ||w||H −1/2 (Γ )
and by using

1 1 1 1
||U ∗ (x, ·)||2H 1 (Ω) = dy + dy
16π 2 |x − y|2 16π 2 |x − y|4
Ω Ω

1 1 1 1 5 |Ω| 1
≤ 2 2
dy+ 2 4
dy ≤ .
4π |x − y0 | π |x − y0 | 4 π |x − y0 |2
2
Ω Ω
The estimate (6.32) follows correspondingly from

∂ 1 yi − xi
u(x) = w(y)dsy .
∂xi 4π |x − y|3
Γ
In the case d = 2 we consider the Taylor expansion

(y − y0 , ȳ − x)
log |y − x| = log |y0 − x| +
|ȳ − x|2
−1/2
with an appropriate ȳ ∈ Ω. Due to w ∈ H∗ (Γ ) we then obtain

1 (y − y0 , ȳ − x)
u(x) = − w(y)dsy ,
2π |ȳ − x|2
Γ
and therefore the estimate (6.31) follows as in the three–dimensional case.

The estimate (6.32) follows in the same way.
For y0 ∈ Ω and R > 2 diam(Ω) we define

BR (y0 ) := x ∈ Rd : |x − y0 | < R .
Then, u(x) = (V w)(x) for x ∈ Ω c is the unique solution of the Dirichlet
boundary value problem
−∆u(x) = 0 for x ∈ BR (y0 )\Ω,

u(x) = γ0ext (V w)(x) = (V w)(x) for x ∈ Γ,
u(x) = (V w)(x) for x ∈ ∂BR (y0 ).
Using Green’s first formula with respect to the bounded domain BR (y0 )\Ω
this gives
aBR (y0 )\Ω (u, v) = −γ1ext u, γ0ext vΓ + γ1int u, γ0int v∂BR (y0 )
where we have used the opposite direction of the exterior normal vector along
Γ . Choosing v = u and using Lemma 6.21 we have

int 1
γ1 u, γ0int u∂BR (y0 ) ≤ c1 (w)c2 (w) dsx ≤ c Rd−4 .
|x − y0 |3
|x−y0 |=R
Hence we can consider the limit R → ∞ to obtain Green’s first formula for
u = V w with respect to the exterior domain as

aΩ c (u, u) := ∇u(x)∇u(x)dx = −γ1ext u, γ0ext uΓ . (6.33)
Ωc
−1/2
Note that in the two–dimensional case the assumption w ∈ H∗ (Γ ) is es-
sential to ensure the above result. In analogy to the estimate (4.17) for the
solution of the interior Dirichlet boundary value problem we find
cext
1 γ1ext u2H −1/2 (Γ ) ≤ aΩ c (u, u). (6.34)
−1/2
Theorem 6.22. Let w ∈ H −1/2 (Γ ) for d = 3 and w ∈ H∗ (Γ ) for d = 2,
respectively. Then there holds
V w, wΓ ≥ cV1 w2H −1/2 (Γ )
with a positive constant cV1 > 0.
Proof. For u = V w we can apply both the interior and exterior Green’s for-
mulae, i.e. (6.28) and (6.33) to obtain
aΩ (u, u) = γ1int u, γ0int uΓ ,

aΩ c (u, u) = −γ1ext u, γ0ext uΓ .
Taking the sum of the above equations we obtain from the jump relation (6.9)
of the single layer potential
aΩ (u, u) + aΩ c (u, u) = [γ1int u − γ1ext u], γ0 uΓ .
The jump relation (6.12) of the conormal derivate of the single layer potential
reads
γ1int u(x) − γ1ext u(x) = w(x) for x ∈ Γ
and therefore we have
aΩ (u, u) + aΩ c (u, u) = V w, wΓ .
Using the inequalities (6.29) and (6.34) this gives

V w, wΓ = aΩ (u, u) + aΩ c (u, u)
≥ cint int 2 ext γ ext u2

1 γ1 uH −1/2 (Γ ) + c1 1 H −1/2 (Γ )
2 3
≥ min{cint ext int 2 ext 2
1 , c1 } γ1 uH −1/2 (Γ ) + γ1 uH −1/2 (Γ ) .
On the other hand, the H −1/2 (Γ ) norm of w can be estimated as
w2H −1/2 (Γ ) = γ1int u − γ1ext u2H −1/2 (Γ )

2 32
≤ γ1int uH −1/2 (Γ ) + γ1ext uH −1/2 (Γ )
2 3
≤ 2 γ1int u2H −1/2 (Γ ) + γ1ext u2H −1/2 (Γ ) .

−1/2
In the two–dimensional case we only have the H∗ (Γ ) ellipticity of the single
layer potential when using the previous theorem. To obtain a more general
result we first consider the following saddle point problem, d = 2, 3, to find
(t, λ) ∈ H −1/2 (Γ ) × R such that
V t, τ Γ − λ1, τ Γ = 0 for all τ ∈ H −1/2 (Γ ),

(6.35)
t, 1Γ = 1.
−1/2
If we consider the ansatz t :=
t + 1/|Γ | for an arbitrary
t ∈ H∗ (Γ ) the
−1/2
second equation is always satisfied. Hence, to find t ∈ H∗ (Γ ) the first
equation reads
1 −1/2
V
t, τ Γ = − V 1, τ Γ for all τ ∈ H∗ (Γ ).
|Γ |
−1/2
The unique solvability of the variational problem follows from the H∗ (Γ )–
ellipticity of the single layer potential V , see Theorem 6.22. The resulting
solution weq := t + 1/|Γ | is denoted as the natural density. By choosing
τ = weq we can finally compute the Lagrange parameter
λ = V weq , weq Γ .
In the three–dimensional case d = 3 it follows from Theorem 6.22 that λ > 0

is strictly positive. In this case the Lagrange parameter λ is called the capacity
of Γ . In the two–dimensional case d = 2 we define by
capΓ := e−2πλ
the logarithmic capacity. For a positive number r ∈ R+ we may define the

parameter dependent fundamental solution
1 1
Ur∗ (x, y) := log r − log |x − y|
2π 2π
which induces an associated boundary integral operator

(Vr w)(x) := Ur∗ (x, y)w(y)dsy for x ∈ Γ
Γ
satisfying
1 1 r
(Vr weq )(x) = log r + λ = log .
2π 2π capΓ
In particular for r = 1 we obtain
1 1
λ := log .
2π capΓ
If the logarithmic capacity capΓ < 1 is strictly less than one, we conclude λ >
0. To ensure capΓ < 1 a sufficient criteria is to assume diam Ω < 1 [81, 157].
This assumption can be always guaranteed when considering a suitable scaling
of the domain Ω ⊂ R2 .
Theorem 6.23. For d = 2 let diam(Ω) < 1 and therefore λ > 0 be satisfied.
The single layer potential V is then H −1/2 (Γ )–elliptic, i.e.,
cV1 w2H −1/2 (Γ )
V w, wΓ ≥ for all w ∈ H −1/2 (Γ ).
Proof. For an arbitrary w ∈ H −1/2 (Γ ) we consider the unique decomposition

−1/2
+ α weq ,
w = w ∈ H∗
w (Γ ), α = w, 1Γ
satisfying
w2H −1/2 (Γ ) = w + αweq 2H −1/2 (Γ )

0 12
≤ w H −1/2 (Γ ) + α weq H −1/2 (Γ )
2 3
≤ 2 w 2H −1/2 (Γ ) + α2 weq 2H −1/2 (Γ )
2 3
≤ 2 max{1, weq 2H −1/2 (Γ ) } w 2H −1/2 (Γ ) + α2 .
On the other hand we have by using V weq , w

Γ =0
V w, wΓ = V (w
+ αweq ), w
+ αweq Γ
= V w,
w Γ + α2 V weq , weq Γ
Γ + 2α V weq , w
2H −1/2 (Γ ) + α2 λ
≥ cV1 w
2 3
≥ min{cV1 , λ} w 2H −1/2 (Γ ) + α2 ,
and therefore the ellipticity estimate follows.

The natural density weq ∈ H −1/2 (Γ ) is a solution of an operator equation
with a constraint,
(V weq )(x) = λ for x ∈ Γ, weq , 1Γ = 1.
By introducing the scaling

eq
weq := λ w
we obtain
1
eq )(x) = 1
(V w for x ∈ Γ, = weq , 1Γ . (6.36)
λ
Instead of the saddle point problem (6.35) we may solve the boundary integral
equation (6.36) to find the natural density weq and afterwards we can compute
the capacity λ by integrating the natural density w eq .
−1/2
The boundary integral operator V : H (Γ ) → H 1/2 (Γ ) is due to (6.8)
−1/2
bounded and H (Γ )–elliptic, see Theorem 6.22 for d = 3 and Theorem 6.23
for d = 2 where we assume diam(Ω) < 1. By the Lax–Milgram theorem (The-
orem 3.4) we therefore conclude the invertibility of the single layer potential
V , i.e. V −1 : H 1/2 (Γ ) → H −1/2 (Γ ) is bounded satisfying (see (3.13))
1
V −1 vH −1/2 (Γ ) ≤ vH 1/2 (Γ ) for all v ∈ H 1/2 (Γ ).
cV1
−1/2
For w ∈ H∗ (Γ ) we have
V w, weq Γ = w, V weq Γ = λw, 1Γ = 0

1/2
and therefore V w ∈ H∗ (Γ ) where
! "
1/2
H∗ (Γ ) := v ∈ H 1/2 (Γ ) : v, weq Γ = 0 .
−1/2 1/2
Thus, V : H∗ (Γ ) → H∗ (Γ ) is an isomorphism.
6.6.2 Ellipticity of the Hypersingular Boundary Integral Operator
Due to (6.17) we have (Du0 )(x) = 0 with the eigensolution u0 (x) ≡ 1 for
x ∈ Γ . Hence we can not ensure the ellipticity of the hypersingular bound-
ary integral operator D on H 1/2 (Γ ). Instead we have to consider a suitable
subspace.
1/2
Theorem 6.24. The hypersingular boundary integral operator D is H∗ (Γ )–
elliptic, i.e.,
2 1/2
Dv, vΓ ≥ cD
1 vH 1/2 (Γ ) for all v ∈ H∗ (Γ ).
1/2
Proof. For v ∈ H∗ (Γ ) we consider the double layer potential
u(x) := −(W v)(x) for x ∈ Ω ∪ Ω c

which is a solution of the homogeneous partial differential equation. The ap-

plication of the trace operators gives
γ0int u(x) = (1 − σ(x))v(x) − (Kv)(x), γ1int u(x) = (Dv)(x) for x ∈ Γ
and
γ0ext u(x) = −σ(x)v(x) − (Kv)(x), γ1ext u(x) = (Dv)(x) for x ∈ Γ.
The function u = −W v is therefore the unique solution of the interior Dirichlet

−∆u(x) = 0 for x ∈ Ω, γ0int u(x) = (1 − σ(x))v(x) − (Kv)(x) for x ∈ Γ
and we have, by applying Green’s first formula (1.5),

∇u(x)∇w(x)dx = γ1int u, γ0int wΓ
Ω
for all w ∈ H 1 (Ω).

For y0 ∈ Ω let BR (y0 ) be a ball of radius R > 2 diam(Ω) which circum-
scribes Ω, Ω ⊂ BR (y0 ). Then, u = −W v is also the unique solution of the
Dirichlet boundary value problem
−∆u(x) = 0 for x ∈ BR (y0 )\Ω,

γ0ext u(x) = −σ(x)v(x) − (Kv)(x) for x ∈ Γ = ∂Ω,
γ0 u(x) = −(W v)(x) for x ∈ ∂BR (y0 )
and the corresponding Green’s first formula reads

∇u(x)∇w(x)dx = −γ1ext u, γ0ext wΓ + γ1 u, γ0 w∂BR (y0 )
BR (y0 )\Ω
for all w ∈ H 1 (BR (y0 )\Ω). For x ∈ Γ we have by definition

1 (y − x, ny )
u(x) = v(y)dsy .
2(d − 1)π |x − y|d
Γ
In particular for x ∈ ∂BR (y0 ) we then obtain the estimates
|u(x)| ≤ c1 (v) R1−d , |∇u(x)| ≤ c2 (v) R−d .
By choosing w = u = −W v and taking the limit R → ∞ we finally obtain

Green’s first formula with respect to the exterior domain,

|∇u(x)|2 dx = −γ1ext u, γ0ext uΓ .
Ωc
By taking the sum of both Green’s formulae with respect to the interior and
to the exterior domain, and considering the jump relations of the boundary in-
tegral operators involved, we obtain for the bilinear form of the hypersingular
boundary integral operator
Dv, vΓ = γ1int u, [γ0int u − γ0ext u]Γ = γ1int u, γ0int uΓ − γ1ext u, γ0ext uΓ

= |∇u(x)|2 dx + |∇u(x)|2 dx = |u|2H 1 (Ω) + |u|2H 1 (Ω c ) .
Ω Ωc
For the exterior domain Ω c we find from the far field behavior of the double
layer potential u(x) = −(W v)(x) as |x| → ∞ the norm equivalence
c1 u2H 1 (Ω c ) ≤ |u|2H 1 (Ω c ) ≤ c2 u2H 1 (Ω c ) .
1/2
For v ∈ H∗ (Γ ), for the natural density weq ∈ H −1/2 (Γ ), V weq = 1, and by
using the symmetry relation (6.26) we further obtain
1 1
γ0int u, weq Γ = ( I − K)v, weq Γ = v, weq Γ − ( I + K)v, weq Γ
2 2
1 −1 −1 1
= −( I + K)v, V 1Γ = −V ( I + K)v, 1Γ
2 2
1 −1 −1 1
= −( I + K )V v, 1Γ = −V v, ( I + K)1Γ = 0
2 2
1/2
and therefore γ0int u ∈ H∗ (Γ ). By using the norm equivalence theorem of
Sobolev (Theorem 2.6) we find
! "1/2
uH∗1 (Ω) := [γ0int u, weq Γ ]2 + ∇u2L2 (Ω)
1/2 1/2
to be an equivalent norm in H 1 (Ω). For v ∈ H∗ (Γ ) we have γ0int u ∈ H∗ (Γ )
and therefore
|u|2H 1 (Ω) = [γ0int u, weq Γ ]2 + ∇u2L2 (Ω) = u2H∗1 (Ω) ≥ c u2H 1 (Ω) .
By using the trace theorem and the jump relation of the double layer potential
we obtain
! "
Dv, vΓ ≥ c u2H 1 (Ω) + u2H 1 (Ω c )
! "
≥ c γ0int u2H 1/2 (Γ ) + γ0ext u2H 1/2 (Γ )
1
≥ c γ0int u − γ0ext u2H 1/2 (Γ ) = cD
2
1 vH 1/2 (Γ )
2
1/2 1/2
for all v ∈ H∗ (Γ ) and therefore the H∗ (Γ )–ellipticity of the hypersingular
boundary integral operator D.
To prove the ellipticity of the hypersingular boundary integral operator

D we have to restrict the functions to a suitable subspace, i.e. orthogonal to
the constants. When considering the orthogonality with respect to different
inner products this gives the ellipticity of the hypersingular boundary integral
operator D with respect to different subspaces.
As in the norm equivalence theorem of Sobolev (cf. Theorem 2.6) we define
!0 12 "1/2
vH∗1/2 (Γ ) := v, weq Γ + |v|2H 1/2 (Γ )
to be an equivalent norm in H 1/2 (Γ ). Here, weq ∈ H −1/2 (Γ ) is the natural

density as defined in (6.36).
Corollary 6.25. The hypersingular boundary integral operator D is H 1/2 (Γ )–

semi–elliptic, i.e.
2
Dv, vΓ ≥ c̄D
1 |v|H 1/2 (Γ ) for all v ∈ H 1/2 (Γ ). (6.37)
1/2
The definition of H∗ (Γ ) involves the natural density weq ∈ H −1/2 (Γ ) as
the unique solution of the boundary integral equation (6.36). From a practi-
cal point of view, this seems not to be very convenient for a computational
realization. Hence we may use a subspace which is induced by a much simpler
inner product. For this we define
! "
1/2
H∗∗ (Γ ) := v ∈ H 1/2 (Γ ) : v, 1Γ = 0 .
1/2
From (6.37) we then have for v ∈ H∗∗ (Γ )
2
Dv, vΓ ≥ c̄D
1 |v|H 1/2 (Γ )
! "
= c̄D
1 |v|2H 1/2 (Γ ) + [v, 1Γ ]2 ≥
cD 2
1 vH 1/2 (Γ ) (6.38)
1/2
the H∗∗ (Γ )–ellipticity of D where we again used the norm equivalence the-
orem of Sobolev (cf. Theorem 2.6).
1/2 (Γ0 ) let
We finally consider an open surface Γ0 ⊂ Γ . For a given v ∈ H
1/2
v ∈ H (Γ ) denote the extension defined by

v(x) for x ∈ Γ0 ,
v(x) =
0 elsewhere.
As in the norm equivalence theorem of Sobolev (cf. Theorem 2.6) we define

! "1/2
wH 1/2 (Γ ),Γ0 := w2L2 (Γ \Γ0 ) + |w|2H 1/2 (Γ )
1/2 (Γ0 )
to be an equivalent norm in H 1/2 (Γ ). Hence we have for v ∈ H
2 3
Dv, vΓ0 = D
v , vΓ ≥ c̄D v |2H 1/2 (Γ ) = c̄D
1 | 1 v 2L2 (Γ \Γ0 ) + |
v |2H 1/2 (Γ )
= c̄D v 2H 1/2 (Γ ),Γ0 ≥ ĉD

1 v 2H 1/2 (Γ ) = ĉD
1
2
1 vH
-1/2 (Γ (6.39)
0)
1/2 (Γ0 )–ellipticity of the hypersingular boundary integral

and therefore the H
operator D.
6.6.3 Steklov–Poincaré Operator
When considering the solution of boundary value problems the interaction of

the Cauchy data γ0int u and γ1int u plays an important role. Let us consider
the system (6.22) of boundary integral equations for a homogeneous partial
differential equation, i.e., f ≡ 0:
& ' & '& '
γ0int u (1 − σ)I − K V γ0int u
= .
γ1int u D σI + K γ1int u
Since the single layer potential V is invertible, we get from the first boundary
integral equation a representation for the Dirichlet to Neumann map,
γ1int u(x) = V −1 (σI + K)γ0int u(x) for x ∈ Γ. (6.40)
The operator
S := V −1 (σI + K) : H 1/2 (Γ ) → H −1/2 (Γ ) (6.41)
is bounded, and S is called Steklov–Poincaré operator. Inserting (6.40) into

the second equation of the Calderón projection this gives
γ1int u(x) = (Dγ0int u)(x) + (σI + K )γ1int u(x)

0 1
= D + (σI + K )V −1 (σI + K) γ0int u(x) for x ∈ Γ. (6.42)
Hence we have obtained a symmetric representation of the Steklov–Poincaré

operator which is equivalent to (6.41),
S := D + (σI + K )V −1 (σI + K) : H 1/2 (Γ ) → H −1/2 (Γ ). (6.43)
With (6.40) and (6.42) we have described the Dirichlet to Neumann map
γ1int u(x) = (Sγ0int )u(x) for x ∈ Γ (6.44)
which maps some given Dirichlet datum γ0int u ∈ H 1/2 (Γ ) to the corresponding
Neumann datum γ1int u ∈ H −1/2 (Γ ) of the harmonic function u ∈ H 1 (Ω)
satisfying Lu = 0.
By using the H 1/2 (Γ )–ellipticity of the inverse single layer potential V −1
we obtain
Sv, vΓ = Dv, vΓ + V −1 (σI + K)v, (σI + K)vΓ ≥ Dv, vΓ (6.45)
for all v ∈ H 1/2 (Γ ). Therefore, the Steklov–Poincaré operator S admits the

same ellipticity estimates as the hypersingular boundary integral operator D.
In particular we have
2 1/2
Sv, vΓ ≥ cD
1 vH 1/2 (Γ ) for all v ∈ H∗ (Γ ) (6.46)
as well as
2 1/2
Sv, vΓ ≥ 1 vH 1/2 (Γ )
cD for all v ∈ H∗∗ (Γ ) (6.47)
while for Γ0 ⊂ Γ we get
Sv, vΓ0 ≥ ĉ1 v2H 1/2 (Γ0 ).

for all v ∈ H (6.48)
-1/2 (Γ 0)
6.6.4 Contraction Estimates of the Double Layer Potential
It is possible to transfer the ellipticity estimates of the single layer potential

V and of the hypersingular boundary integral operator D to the double layer
potential σI + K : H 1/2 (Γ ) → H 1/2 (Γ ), see [145]. Since the single layer
potential V : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded and H −1/2 (Γ )–elliptic, we
may define
.
uV −1 := V −1 u, uΓ for all u ∈ H 1/2 (Γ )
to be an equivalent norm in H 1/2 (Γ ).

1/2
Theorem 6.26. For u ∈ H∗ (Γ ) we have
(1 − cK ) uV −1 ≤ (σI + K)uV −1 ≤ cK uV −1 (6.49)
with 8
1 1
cK = + − cV1 cD
1 <1 (6.50)
2 4
where cV1 and cD
1 are the ellipticity constants of the single layer potential V
and of the hypersingular boundary integral operator D, respectively.
Proof. Using the symmetric representation (6.43) of the Steklov–Poincaré op-

erator S we have
(σI + K)u2V −1 = V −1 (σI + K)u, (σI + K)uΓ

= Su, uΓ − Du, uΓ .
Let J : H −1/2 (Γ ) → H 1/2 (Γ ) be the Riesz map which is defined via
Jw, vH 1/2 (Γ ) = w, vΓ for all v ∈ H 1/2 (Γ ).

Then, A := JV −1 : H 1/2 (Γ ) → H 1/2 (Γ ) is self–adjoint and H 1/2 (Γ )–elliptic.

Using the first representation (6.41) of the Steklov–Poincaré operator S and
considering the splitting A = A1/2 A1/2 we conclude the inequality
Su, uΓ = V −1 (σI + K)u, uΓ

= JV −1 (σI + K)u, uH 1/2 (Γ )
= A1/2 (σI + K)u, A1/2 uH 1/2 (Γ )
≤ A1/2 (σI + K)uH 1/2 (Γ ) A1/2 uH 1/2 (Γ )
and with
A1/2 v2H 1/2 (Γ ) = A1/2 v, A1/2 vH 1/2 (Γ )

= JV −1 v, vH 1/2 (Γ )
= V −1 v, vΓ
= v2V −1
it follows that
Su, uΓ ≤ (σI + K)uV −1 uV −1 .
1/2
Since the hypersingular integral operator D is elliptic for u ∈ H∗ (Γ ) we
find from the mapping properties of the inverse single layer potential V −1 the
lower estimate
2 −1 2
Du, uΓ ≥ cD
1 uH 1/2 (Γ ) ≥ c1 c1 V
D V
1 c1 uV −1 .
u, uΓ = cD V
Hence we have obtained
(σI + K)u2V −1 = Su, uΓ − Du, uΓ

2
≤ (σI + K)uV −1 uV −1 − cV1 cD
1 uV −1 .
Denoting
a := (σI + K)uV −1 ≥ 0, b := uV −1 > 0
we conclude 9 a :2 a
− 1 ≤ 0
+ cV1 cD
b b
which is equivalent to
8 8
1 1 a 1 1
− − c1 c1 ≤
V D ≤ + − cV1 cD
1
2 4 b 2 4
and therefore to the assertion.
The contraction property of σI + K, in particular the upper estimate in
(6.49), can be extended to hold in H 1/2 (Γ ).
Corollary 6.27. For u ∈ H 1/2 (Γ ) there holds
(σI + K)uV −1 ≤ cK uV −1 (6.51)
where the contraction rate cK < 1 is given as in (6.50).

Proof. For an arbitrary u ∈ H 1/2 (Γ ) we can write
u, weq Γ
+
u = u u0
1, weq Γ
1/2
∈ H∗ (Γ ) and u0 ≡ 1. Due to (σI + K)u0 = 0 we have by using
where u
Theorem 6.26
(σI + K)uV −1 = (σI + K)

uV −1 ≤ cK
uV −1 .
On the other hand,
[u, weq Γ ]2
u2V −1 =
u2V −1 + u2V −1
≥
1, weq Γ
which implies the contraction estimate (6.51).

Note that a similar result as in Theorem 6.26 can be shown for the shifted
operator(1 − σ)I − K.
1/2
Corollary 6.28. For v ∈ H∗ (Γ ) there holds
(1 − cK ) vV −1 ≤ ([1 − σ]I − K)vV −1 ≤ cK vV −1 (6.52)

Proof. By using both the triangle inequality and the contraction estimate of
σI + K we obtain with
vV −1 = ([1 − σ]I − K)v + (σI + K)vV −1

≤ ([1 − σ]I − K)vV −1 + (σI + K)vV −1
≤ ([1 − σ]I − K)vV −1 + cK vV −1
the lower estimate in (6.52). Moreover, using both representations (6.41) and
(6.43) of the Steklov–Poincaré operator S, we conclude
((1 − σ)I − K)v2V −1 = [I − (σI + K)]v2V −1

= v2V −1 + (σI + K)v2V −1 − 2V −1 (σI + K)v, vΓ
= v2V −1 + (σI + K)v2V −1 − 2Sv, vΓ
= v2V −1 − (σI + K)v2V −1 − 2Dv, vΓ
0 1
≤ 1 − (1 − cK )2 − 2cV1 cD
1 v2V −1 = c2K v2V −1 .
This gives the upper estimate in (6.52).

−1/2 −1/2
Let H∗ (Γ ) be the subspace as defined in (6.30). Due to V : H∗ (Γ ) →
1/2
H∗ (Γ ) we can transfer the estimates (6.49) of the double layer potential
σI + K : H 1/2 (Γ ) → H 1/2 (Γ ) immediately to the adjoint double layer poten-
tial σI + K : H −1/2 (Γ ) → H −1/2 (Γ ).
−1/2
Corollary 6.29. For the adjoint double layer potential and for w ∈ H∗ (Γ )
there holds
(1 − cK ) wV ≤ (σI + K )wV ≤ cK wV (6.53)
where the contraction rate cK < 1 is given as in (6.50) and · V is the norm
which is induced by the single layer potential V .
−1/2 1/2
Proof. For w ∈ H∗ (Γ ) there exists a uniquely determined v ∈ H∗ (Γ )
satisfying v = V w or w = V −1 v. Using the symmetry property (6.26) we first
have
(σI + K )w2V = V (σI + K )V −1 v, (σI + K )V −1 vΓ

= V −1 (σI + K)v, (σI + K)vΓ = (σI + K)v2V −1 ,
as well as
w2V = V w, wΓ = V −1 v, vΓ = v2V −1 .
Therefore, (6.53) is equivalent to (6.49).

As in Corollary 6.27 we can extend the contraction property of σI + K
to H −1/2 (Γ ).
Corollary 6.30. For w ∈ H −1/2 (Γ ) there holds the contraction estimate
(σI + K )wV ≤ cK wV . (6.54)
To prove related properties of the shifted adjoint double layer potential

(1 − σ)I − K again we need to bear in mind the correct subspaces.
−1/2
Corollary 6.31. For w ∈ H∗ (Γ ) we have
(1 − cK ) wV ≤ ((1 − σ)I − K )wV ≤ cK wV (6.55)
6.6.5 Mapping Properties
All mapping properties of boundary integral operators considered up to now

0 : H
are based on the mapping properties of the Newton potential N −1 (Ω) →
1
H (Ω), and on the application of trace theorems and on duality arguments.
But even for Lipschitz domains more general results hold.
0 : H
Theorem 6.32. The Newton potential N s (Ω) → H s+2 (Ω) is a continu-
ous map for all s ∈ [−2, 0], i.e.
0 f H s+2 (Ω) ≤ c f -s
N s (Ω).
for all f ∈ H
H (Ω)
Proof. Let s ∈ (−1, 0] and consider f to be the extension of f ∈ H s (Ω) as

defined in (6.2). Then,
f, vRd f, vΩ

fH s (Rd ) = sup ≤ sup = f H s (Ω) ,
−s
0=v∈H (R )
d v H −s (R d ) 0 = v∈H −s (Ω) vH −s (Ω)
and the assertion follows as in the proof of Theorem 6.1, i.e.

0 f H s+2 (Ω) ≤ c f -s
N s (Ω).
for all f ∈ H
H (Ω)
0 is self–adjoint, for s ∈ [−2, −1) we obtain by

Since the Newton potential N
duality
0 f, gΩ
N 0 gΩ
f, N
0 f H s+2 (Ω) =
N sup = sup
-−2−s (Ω) gH
0=g∈H -−2−s (Ω) -−2−s (Ω) gH
0=g∈H -−2−s (Ω)
0 gH −s (Ω)
N
≤ f H
-s (Ω) sup ≤ c f H
-s (Ω) .

-−2−s (Ω) gH
0=g∈H -−2−s (Ω)
By the application of Theorem 6.32 we can deduce corresponding mapping

properties for the single layer potential V as defined in (6.7) and for the
boundary integral operator V := γ0int V by considering the trace of V .
1 1
Theorem 6.33. The single layer potential V : H − 2 +s (Γ ) → H 2 +s (Γ ) is
bounded for |s| < 12 , i.e.,
V wH 1/2+s (Γ ) ≤ c wH −1/2+s (Γ )
for all w ∈ H −1/2+s (Γ ).
Proof. For ϕ ∈ C ∞ (Ω) we first consider

∗
V w, ϕΩ = ϕ(x) U (x, y)w(y)dsy dx = w(y) U ∗ (x, y)ϕ(x)dxdsy
Ω Γ Γ Ω
0 ϕΓ ≤ wH −1/2+s (Γ ) γ int N
= w, γ0int N 0 ϕH 1/2−s (Γ )
0
0 ϕH 1−s (Ω)
≤ cT wH −1/2+s (Γ ) N
1
where the application of the trace theorem requires 2 − s > 0, see Theorem
2.21. With Theorem 6.32 we then obtain
V w, ϕΩ ≤ c wH −1/2+s (Γ ) ϕH

-−1−s (Ω) for all ϕ ∈ C ∞ (Ω).
By using a density argument we conclude V w ∈ H 1+s (Ω). Taking the trace

this gives V w := γ0int V w ∈ H 1/2+s (Γ ) when assuming 12 + s > 0.

In the case of a Lipschitz domain Ω we can prove as in Theorem 6.33
corresponding mapping properties for all boundary integral operators.
Theorem 6.34. [44] Let Γ := ∂Ω be the boundary of a Lipschitz domain Ω.

Then, the boundary integral operators
V : H −1/2+s (Γ ) → H 1/2+s (Γ ),
K : H 1/2+s (Γ ) → H 1/2+s (Γ ),
K : H −1/2+s (Γ ) → H −1/2+s (Γ ),
D : H 1/2+s (Γ ) → H −1/2+s (Γ )
are bounded for all s ∈ [− 12 , 12 ].
Proof. For the single layer potential V and for |s| < 12 the assertion was
already shown in Theorem 6.33. This remains true for |s| = 12 [152], see also
the discussion in [103].
For all other boundary integral operators the assertion follows from the
mapping properties of the conormal derivative operator.
First we consider the adjoint double layer potential K . Recall that the sin-
gle layer potential u(x) = (V w)(x), x ∈ Ω, is a solution of the homoge-
neous partial differential equation with Dirichlet data γ0int u(x) = (V w)(x) for
x ∈ Γ . By using Theorem 4.6 and the continuity of the single layer potential
V : L2 (Γ ) → H 1 (Γ ) we obtain
γ1int uL2 (Γ ) ≤ c V wH 1 (Γ ) ≤
c wL2 (Γ )
and therefore the continuity of γ1int V = σI + K : L2 (Γ ) → L2 (Γ ). On the

other hand we have by duality
γ1int u, ϕΓ
γ1int uH −1 (Γ ) = sup .
0=ϕ∈H 1 (Γ ) ϕH 1 (Γ )
For an arbitrary ϕ ∈ H 1 (Γ ) let v ∈ H 3/2 (Ω) be the unique solution of the

Dirichlet boundary value problem Lv(x) = 0 for x ∈ Ω and γ0int v(x) = ϕ(x)
for x ∈ Γ . By using Theorem 4.6 this gives
γ1int vL2 (Γ ) ≤ c γ0int vH 1 (Γ ) = c ϕH 1 (Γ ) .
Since both u = V w and v are solutions of a homogeneous partial differential

equation, we obtain by applying Green’s first formula (1.5) twice and taking
into account the symmetry of the bilinear form
γ1int u, ϕΓ = a(u, v) = a(v, u) = γ1int v, γ0int uΓ

≤ γ1int vL2 (Γ ) γ0int uL2 (Γ ) ≤ c ϕH 1 (Γ ) V wL2 (Γ ) .
From the continuity of the single layer potential V : H −1 (Γ ) → L2 (Γ ) we
now conclude
γ1int uH −1 (Γ ) ≤ c V wL2 (Γ ) ≤
c wH −1 (Γ )
and therefore the continuity of γ1int V = σI + K : H −1 (Γ ) → H −1 (Γ ). Using

an interpolation argument we obtain K : H −1/2+s (Γ ) → H −1/2+s (Γ ) for all
|s| ≤ 12 . Due to
Kv, wΓ
KvH 1/2+s (Γ ) = sup
0=w∈H −1/2−s (Γ ) wH −1/2−s (Γ )
v, K wΓ
= sup
0=w∈H −1/2−s (Γ ) wH −1/2−s (Γ )
K wH −1/2−s (Γ )
= vH 1/2+s (Γ ) sup
0=w∈H −1/2−s (Γ ) wH −1/2−s (Γ )
≤ c vH 1/2+s (Γ )
we immediately conclude K : H 1/2+s (Γ ) → H 1/2+s (Γ ) for |s| ≤ 12 .

It remains to prove the assertion for the hypersingular boundary integral
operator D. The double layer potential u(x) = (W v)(x), x ∈ Ω, yields, by the
application of Theorem 4.6,
DvL2 (Γ ) = γ1int uL2 (Γ ) ≤ c γ0int uH 1 (Γ ) = c ([σ − 1]I + K)vH 1 (Γ )

and therefore D : H 1 (Γ ) → L2 (Γ ). Again, using duality and interpolation
arguments completes the proof.
If the boundary Γ = ∂Ω of the bounded domain Ω ⊂ Rd is piecewise
smooth, Theorem 6.34 remains true for larger values of |s|. For example, if
Ω ⊂ R2 is polygonal bounded with J corner points and associated interior
angles αj we may define

π π
σ0 := min min , .
j=1,...,J αj 2π − αj
Then, Theorem 6.34 holds for all |s| < σ0 [45]. If the boundary Γ is C ∞ , then
Theorem 6.34 remains true for all s ∈ R.

All mapping properties of boundary integral operators as shown above for the
model problem of the Laplace equation can be transfered to general second
order partial differential equations, when a fundamental solution is known. In

what follows we will consider the system of linear elastostatics which reads
for d = 2, 3 and x ∈ Ω ⊂ Rd as
E E
− ∆u(x) − grad div u(x) = f (x). (6.56)
2(1 + ν) 2(1 + ν)(1 − 2ν)
The associated fundamental solution is the Kelvin tensor (5.9)

∗ 1 1 1+ν (xi − yi )(xj − yj )
Uij (x, y) = (3 − 4ν)E(x, y)δij +
4(d − 1)π E 1 − ν |x − y|d
for i, j = 1, . . . , d where
⎧
⎨ − log |x − y|
⎪ for d = 2,
E(x, y) = 1
⎪
⎩ for d = 3.
|x − y|
For the components ui of the solution there holds the representation formula
∈ Ω, i = 1, . . . , d,
(5.10) (Somigliana identity), x

d
d
∗
ui (
x) = Uij (
x, y)tj (y)dsy − Tij∗ (
x, y)uj (y)dsy
Γ j=1 Γ j=1

d
∗
+ Uij (
x, y)fj (y)dy. (6.57)
Ω j=1
As in (6.4) we define the Newton potential

d
0 f )i (
(N x) = ∗
Uij (
x, y)fj (y)dy ∈ Ω, i = 1, . . . , d
for x
Ω j=1
which is a generalized solution of (6.56). Moreover, as in Theorem 6.1,
0 f [H 1 (Ω)]d ≤ c f -−1
N .
[H (Ω)]d
By taking the interior traces of ui ,
0 f )i (x) =
(N0 f )i (x) := γ0int (N lim 0 f )i (
(N x) for i = 1, . . . , d,
x→x∈Γ
Ω
this defines a linear and bounded operator
0 : [H
N0 := γ0int N −1 (Ω)]d → [H 1/2 (Γ )]d .
In addition, by applying the boundary stress operator (1.23),


d
0 f )i (x) =
(γ1int N lim 0 f , x
σij (N )nj (x) for i = 1, . . . , d,
x→x∈Γ
Ω
j=1
we introduce a second linear and bounded operator

0 : [H
N1 := γ1int N −1 (Ω)]d → [H −1/2 (Γ )]d .
∈ Ω ∪ Ω c the single layer potential

For x

d
(V w)i (
x) = ∗
Uij (
x, y)wj (y)dsy for i = 1, . . . , d
Γ j=1
is a solution of the homogeneous system of linear elasticity (6.56) with f = 0.

Hence we have
V : [H 1/2 (Γ )]d → [H 1 (Ω)]d .
When considering the interior and exterior traces of V this defines a bounded
linear operator
V := γ0int V = γ0ext V : [H −1/2 (Γ )]d → [H 1/2 (Γ )]d
with a representation as a weakly singular surface integral,

d
∗
(V w)i (x) = Uij (x, y)wj (y)dsy for x ∈ Γ, i = 1, . . . , d. (6.58)
Γ j=1
If w ∈ [H −1/2 (Γ )]d is given, the single layer potential V w ∈ [H 1 (Ω)]d is a

solution of the homogeneous system (6.56) of linear elastostatics. Then the
application of the interior boundary stress operator (1.23),

d
(γ1int V w)i (x) = lim σij (V w, x
)nj (x) for i = 1, . . . , d,
x→x∈Γ
Ω
j=1
defines a bounded linear operator
γ1int V : [H −1/2 (Γ )]d → [H −1/2 (Γ )]d
with the representation

1
(γ1int V w)i (x) = wi (x) + (K w)i (x) for almost all x ∈ Γ, i = 1, . . . , d,
2
where

d
d
(K w)k (x) = lim σk (U ∗j (x, y), x)n (x)wj (y)dsy
ε→0
j=1 =1
is the adjoint double layer potential, k = 1, . . . , d. For simplicity we may

assume that x ∈ Γ is on a smooth part of the surface, in particular we do
not consider the case when x ∈ Γ is either a corner point or on an edge.
Correspondingly, the application of the exterior boundary stress operator gives
1
(γ1ext V w)i (x) = − wi (x) + (K w)i (x) for almost all x ∈ Γ, i = 1, . . . , d.
2
Hence we obtain the jump relation for the boundary stress of the single layer
potential as
[γ1 V w] := (γ1ext V w)(x) − (γ1int V w)(x) = −w(x) for x ∈ Γ

in the sense of [H −1/2 (Γ )]d .
As for the Laplace operator the far field behavior of the single layer poten-
tial V is essential when investigating the ellipticity of the single layer potential
V . The approach as considered for the Laplace equation can be applied as well
for the system of linear elastostatics. Note that the related subspace is now
induced by the rigid body motions (translations). Hence, for d = 2, we define
! "
−1/2
[H+ (Γ )]2 := w ∈ [H −1/2 (Γ )]2 : wi , 1Γ = 0 for i = 1, 2 .
Lemma 6.35. For y0 ∈ Ω and x ∈ R3 let |x − y0 | > 2 diam(Ω) be satisfied.

−1/2
Assume w ∈ [H −1/2 (Γ )]3 for d = 3 and w ∈ [H+ (Γ )]2 for d = 2. For
u(x) = (V w)(x) we then have
1
|ui (x)| ≤ c , i = 1, . . . , d.
|x − y0 |
Proof. Let d = 3. Due to

1 1 1+ν δij (xi − yi )(xj − yj )
|ui (x)| = (3 − 4ν) + w (y)ds

4π E 1 − ν |x − y| |x − y| 3 j y

Γ

1 1 1+ν δij 1
≤ (3 − 4ν) + |wj (y)|dsy
4π E 1 − ν |x − y| |x − y|
Γ
we obtain the assertion as in the proof of Lemma 6.21. For d = 2 we consider

the Taylor expansion of the fundamental solution to conclude the result as in
the proof of Lemma 6.21.
As for the single layer potential of the Laplace operator (Theorem 6.22)
we now can prove the following ellipticity result.
−1/2
Theorem 6.36. Assume w ∈ [H −1/2 (Γ )]d for d = 3 and w ∈ [H+ (Γ )]d
for d = 2. Then we have
V w, wΓ ≥ cV1 w2[H −1/2 (Γ )]d
with a positive constant cV1 .

Proof. The ellipticity estimate follows as is the proof of Theorem 6.22 by using
Lemma 4.19.
To prove the [H −1/2 (Γ )]2 –ellipticity of the two–dimensional single layer
potential V we first introduce the generalized fundamental solution

1 1 1+ν (xi − yi )(xj − yj )
α
Uij (x, y) = (4ν − 3) log(α|x − y|)δij +
4π E 1 − ν |x − y|2
for i, j = 1, 2 which depends on a real parameter α ∈ R+ , and we consider the
corresponding single layer potential Vα : [H −1/2 (Γ )]2 → [H −1/2 (Γ )]2 . Note
that this approach corresponds to some scaling of the computational domain
Ω ⊂ R2 and its boundary Γ , respectively.
−1/2
For w ∈ [H+ (Γ )]2 we have by using Theorem 6.36 the ellipticity esti-
mate
Vα w, wΓ = V w, wΓ ≥ cV1 w2[H −1/2 (Γ )]2 .
The further approach now corresponds to the case of the scalar single layer
potential of the Laplace operator [142] to find (w1 , λ1 ) ∈ [H −1/2 (Γ )]2 × R2 as
the solution of the saddle point problem
Vα w1 , τ Γ − λ11 1, τ1 Γ − λ12 1, τ2 Γ = 0
w11 , 1Γ =1
1
w2 , 1Γ =0
to be satisfied for all τ ∈ [H −1/2 (Γ )]2 . By introducing w11 := w

11 + 1/|Γ | and
−1/2
w21 := w21 it remains to find w1 ∈ [H+ (Γ )]2 as the unique solution of the
variational problem
1 −1/2
1 , τ Γ = −
Vα w Vα (1, 0) , τ Γ for all τ ∈ [H+ (Γ )]2 .
|Γ |
−1/2
When w1 ∈ [H+ (Γ )]2 and therefore w1 ∈ [H −1/2 (Γ )]2 is known we can
compute
λ11 = Vα w1 , w1 Γ .
In the same way we find (w2 , λ2 ) ∈ [H −1/2 (Γ )]2 × R2 as the solution of the
saddle point problem
Vα w2 , τ Γ − λ21 1, τ1 Γ − λ22 1, τ2 Γ = 0
w12 , 1Γ =0
w22 , 1Γ =1
to be satisfied for all τ ∈ [H −1/2 (Γ )]2 . Moreover, we obtain
λ22 = Vα w2 , w2 Γ ,
as well as
λ12 = λ21 = Vα w1 , w2 Γ .
Lemma 6.37. For the Lagrange multiplier λ1i (i = 1, 2) we have the repre-
sentation
1 1 1+ν
λii = V wi , wi Γ + (4ν − 3) log α,
4π E 1 − ν
while the Lagrange multiplier λ12 = λ21 is independent of α ∈ R+ ,
λ12 = λ21 = V w1 , w2 Γ .
Proof. For i = 1, a direct computation gives, by splitting the fundamental
solution log(α|x − y|),
λ11 = Vα w1 , w1 Γ
2
1 1 1+ν
= (4ν − 3) log(α|x − y|)wi1 (y)wi1 (x)dsx dsy
4π E 1 − ν i=1
Γ Γ
2
1 1 1+ν (xi − yi )(xj − yj ) 1
+ 2
wi (y)wj1 (x)dsx dsy
4π E 1 − ν i,j=1
|x − y|
Γ Γ
2
1 1 1+ν
= (4ν − 3) log |x − y|wi1 (y)wi1 (x)dsx dsy
4π E 1 − ν i=1Γ Γ
2
1 1 1+ν (xi − yi )(xj − yj ) 1
+ 2
wi (y)wj1 (x)dsx dsy
4π E 1 − ν i,j=1
|x − y|
Γ Γ
2
1 1 1+ν 0 1 12
+ (4ν − 3) log α wi , 1Γ
4π E 1 − ν i=1
1 1 1+ν
= V1 w1 , w1 Γ + (4ν − 3) log α
4π E 1 − ν
due to w11 , 1Γ = 1 and w21 , 1Γ = 0. For λ22 the assertion follows in the same
way. Finally, for λ21 = λ12 we have
λ21 = Vα w1 , w2 Γ
2
1 1 1+ν
= V1 w1 , w2 Γ + (4ν − 3) log α wi1 , 1Γ wi2 , 1Γ
4π E 1 − ν i=1
= V1 w1 , w2 Γ
due to w12 , 1Γ = w21 , 1Γ = 0.

Hence we can choose the scaling parameter α ∈ R+ such that
min{λ11 , λ22 } ≥ 2 |λ12 | (6.59)
is satisfied. An arbitrary given w ∈ [H −1/2 (Γ )]2 can be written as
+ α1 w1 + α2 w2 ,
w = w αi = wi , 1Γ (i = 1, 2) (6.60)
−1/2
∈ [H+
where w (Γ )]2 .
Theorem 6.38. Let the scaling parameter α ∈ R+ be chosen such that (6.59)
is satisfied. Then the single layer potential Vα is [H −1/2 (Γ )]2 –elliptic, i.e.
cV1 w2[H −1/2 (Γ )]2
Vα w, wΓ ≥ for all w ∈ [H −1/2 (Γ )]2 .
Proof. For an arbitrary w ∈ [H −1/2 (Γ )]2 we consider the splitting (6.60).

By using the triangle inequality as well as the Cauchy–Schwarz inequality we
obtain
w2[H −1/2 (Γ )]2 = w + α1 w1 + α2 w2 2[H −1/2 (Γ )]2
2 32
≤ w [H −1/2 (Γ )]2 + |α1 | w1 2[H −1/2 (Γ )]2 + |α2 | w2 2[H −1/2 (Γ )]2
2 3
≤ 3 w 2[H −1/2 (Γ )]2 + α12 w1 2[H −1/2 (Γ )]2 + α22 w2 2[H −1/2 (Γ )]2
! "2 3
≤ 3 max 1, w1 2[H −1/2 (Γ )]2 , w2 2[H−1/2(Γ )]2 w 2[H −1/2 (Γ )]2 + α12 + α22 .
Moreover, by the construction of w1 and w2 we have

+ α1 w1 + α2 w2 ], w
Vα w, wΓ = Vα [w + α1 w1 + α2 w2 Γ
= Vα w, Γ + α12 Vα w1 , w1 Γ + α22 Vα w2 , w2 Γ

w
+2α1 Vα w1 , w
Γ + 2α2 Vα w2 , w
Γ + 2α1 α2 Vα w1 , w2 Γ
= Vα w, Γ + α12 λ11 + α22 λ22 + 2α1 α2 λ21 .

w
−1/2
From the [H+ (Γ )]2 –ellipticity of Vα and by using the scaling condition
(6.59) we finally get
2[H −1/2 (Γ )]2 + α12 λ11 + α22 λ22 − 2|α1 | |α2 | |λ21 |
Vα w, wΓ ≥ cV1 w
0 1
2[H −1/2 (Γ )]2 + min{λ11 , λ22 } α12 + α22 − |α1 | |α2 |
≥ cV1 w
1 0 1
2[H −1/2 (Γ )]2 + min{λ11 , λ22 } α12 + α22
≥ cV1 w
2
2 3
1 1
≥ min cV1 , λ11 , λ22 2[H −1/2 (Γ )]2 + α12 + α22 .
w

2 2
Therefore, the single layer potential V : [H −1/2 (Γ )]d → [H 1/2 (Γ )]d is

bounded and [H −1/2 (Γ )]d –elliptic, where for d = 2 we have to assume a
suitable scaling of the domain Ω, see (6.59). By using the Lax–Milgram lemma
(Theorem 3.4) we therefore conclude the existence of the inverse operator
V −1 : [H 1/2 (Γ )]d → [H −1/2 (Γ )]d .
By R we denote the set of rigid body motions, i.e. (1.36) for d = 2 and
(1.29) for d = 3, respectively. Define
! "
−1/2
[H∗ (Γ )]d := w ∈ [H −1/2 (Γ )]d : w, v k Γ = 0 for v k ∈ R ,
and
! "
1/2
[H∗ (Γ )]d := v ∈ [H 1/2 (Γ )]d : V −1 v, v k Γ = 0 for v k ∈ R .
−1/2 1/2
Obviously, V : [H∗ (Γ )]d → [H∗ (Γ )]d is an isomorphism.
∈ Ω ∪ Ω we define by
For x c

d
(W v)i (
x) := Tij∗ (
x, y)uj (y)dsy , i = 1, . . . , d,
Γ j=1
the double layer potential of linear elastostatics satisfying
W : [H 1/2 (Γ )]d → [H 1 (Γ )]d .
The application of the interior trace operator defines a bounded linear oper-
ator
γ0int W : [H 1/2 (Γ )]d → [H 1/2 (Γ )]d
with the representation
1
(γ0int W v)i (x) = − vi (x) + (Kv)i (x) for almost all x ∈ Γ, i = 1, . . . , d,
2
(6.61)
where

d
(Kv)i (x) := lim Tij∗ (
x, y)uj (y)dsy , i = 1, . . . , d
ε→0
j=1
is the double layer potential. Correspondingly, the application of the exterior

trace operator gives
1
(γ0ext W v)i (x) = vi (x) + (Kv)i (x) for almost all x ∈ Γ, i = 1, . . . , d.
2
Hence, we obtain the jump relation of the double layer potential as
[γ0 W v] = γ0ext (W v)(x) − γ0int (W v)(x) = v(x) for x ∈ Γ.
From the Somigliana identity (6.57) we get for Ω x

→ x ∈ Γ by using (6.58)
and (6.61) the boundary integral equation

1
(V t)(x) = I + K u(x) − (N0 f )(x) for almost all x ∈ Γ. (6.62)
2
Inserting the rigid body motions (1.36) for d = 2 and (1.29) for d = 3 this
gives
1
I + K v k (x) = 0 for x ∈ Γ and v k ∈ R.
2
The application of the interior boundary stress operator γ1int on the double
layer potential W v defines a bounded linear operator
γ1int W = γ1ext W : [H 1/2 (Γ )]d → [H −1/2 (Γ )]d .
As in the case of the Laplace operator we denote by D := −γ1int W the hyper-

singular boundary integral operator.
When applying the boundary stress operator γ1int on the Somigliana iden-
tity (6.57) this gives the hypersingular boundary integral equation

1
(Du)(x) = I − K t(x) − (N1 f )(x) for x ∈ Γ (6.63)
2
in the sense of [H −1/2 (Γ )]d .

As in (6.22) we can write both boundary integral equations (6.62) and
(6.63) as a system with the Calderón projection (6.23). Note that the projec-
tion property of the Calderón projection (Lemma 6.18) as well as all relations
of Corollary 6.19 remain valid as for the scalar Laplace equation.
Analogous to the Laplace operator we can rewrite the bilinear form of the
hypersingular boundary integral operator by using integration by parts as a
sum of weakly singular bilinear forms. In particular for d = 2 we have the
representation [107]
2

Du, vΓ = curlΓ vj (x) Gij (x, y)curlΓ ui (y)dsy dsx
i,j=1 Γ Γ
for all u, v ∈ [H 1/2 (Γ ) ∩ C(Γ )]2 where

1 E (xi − yi )(xj − yj )
Gij (x, y) = − log |x − y| δij + , i, j = 1, 2.
4π 1 − ν 2 |x − y|2
Here, curlΓ denotes the derivative with respect to the arc length. Note that the
kernel functions Gij (x, y) correspond, up to constants, to the kernel functions
of the Kelvin fundamental solution (5.9).
In the three–dimensional case d = 3 and i, j = 1, . . . , 3 we define
∂ ∂
Mij (∂x , n(x)) := nj (x) − ni (x)
∂xi ∂xj
and
∂
:= M32 (∂x , n(x)),
∂S1 (x)
∂
:= M13 (∂x , n(x)),
∂S2 (x)
∂
:= M21 (∂x , n(x)).
∂S3 (x)
The bilinear form of the hypersingular boundary integral operator D can then
be written as [75]
& 3 '
µ 1 ∂ ∂
Du, vΓ = u(y) · v(x) dsy dsx
4π |x − y| ∂Sk (y) ∂Sk (x)
Γ Γ k=1

µ I 2 ∗
+ (M (∂x , n(x))v(x)) −4µ U (x, y) M (∂y , n(y))u(y)dsy dsx
2π |x − y|
ΓΓ
3

µ 1
+ Mkj (∂x , n(x))vi (x) Mki (∂y , n(y))vj (y)dsy dsx . (6.64)
4π |x − y|
Γ Γ i,j,k=1
Hence we can express the bilinear form of the hypersingular boundary integral
operator D by components of the single layer potential V only.
Moreover, for d = 3 there holds a related representation of the double layer
potential K, see [89],

1 ∂ 1 1 1
(Ku)(x) = u(y)dsy − M (∂y , n(y))u(y)dsy
4π ∂n(y) |x − y| 4π |x − y|
Γ Γ
E
+ (V (M (∂· , n(·))u(·)))(x)
1+ν
where the evaluation of the Laplace single and double layer potentials has to
be taken componentwise.
Hence we can reduce all boundary integral operators of linear elastostatics
to the single and double layer potentials of the Laplace equation. These rela-
tions can be used when considering the Galerkin discretization of boundary
integral equations, where only weakly singular surface integrals have to be
computed.
Inserting the rigid body motions (1.36) for d = 2 and (1.29) for d = 3 into
the representation formula (6.57) this gives
x) = −(W v k )(
v k ( x) ∈ Ω and v k ∈ R.
for x
The application of the boundary stress operator γ1int yields
(Dv k )(x) = 0 for x ∈ Γ and v k ∈ R.

1/2
As in Theorem 6.24 we can prove the [H∗ (Γ )]d –ellipticity of the hypersin-
gular boundary integral operator D,
2 1/2
Dv, vΓ ≥ cD
1 v[H 1/2 (Γ )]d for all v ∈ [H∗ (Γ )]d .
In addition, the ellipticity of the hypersingular boundary integral operator D

can be formulated also in the subspace
6.8 Stokes System 165
! "
1/2
[H∗∗ (Γ )]d := v ∈ [H 1/2 (Γ )]d : v, v k Γ = 0 for v k ∈ R
of functions which are orthogonal to the rigid body motions. Then we have,
as in (6.38),
2 1/2
Dv, vΓ ≥ 1 v[H 1/2 (Γ )]d
cD for all v ∈ [H∗∗ (Γ )]d .
As in Subsection 6.6.3 we can define the Dirichlet to Neumann map, which

relates given boundary displacements to the associated boundary stresses via
the Steklov–Poincaré operator. Moreover, all results on the contraction prop-
erty of the double layer potential (see Subsection 6.6.4) as well as all mapping
properties of boundary integral operators (see Subsection 6.6.5) remain valid
for the system of linear elastostatics.
6.8 Stokes System

Now we consider the homogeneous Stokes system (1.38) where we assume
µ = 1 for simplicity, i.e.,
−∆u(x) + ∇p(x) = 0, div u(x) = 0 for x ∈ Ω.
Since the fundamental solution of the Stokes system coincides with the Kelvin
fundamental solution of linear elastostatics when considering ν = 12 and E = 3
as material parameters, we can write the representation formula (6.57) and
all related boundary integral operators of linear elastostatics for ν = 12 and
E = 3 to obtain the Stokes case. However, since the analysis of the mapping
properties of all boundary integral operators of linear elastostatics assumes
ν ∈ (0, 12 ) we can not transfer the boundedness and ellipticity estimates from
linear elastostatics to the Stokes system. These results will be shown by con-
sidering the Stokes single layer potential which is also of interest for the case
of almost incompressible linear elasticity (ν = 12 ) [141].
Let Ω ⊂ Rd be a simple connected domain with boundary Γ = ∂Ω. The
single layer potential V : [H −1/2 (Γ )]d → [H 1 (Ω)]d induces a function

d
x) := (V w)i (
ui ( x) = ∗
Uij (
x, y)wj (y)dsy ∈ Ω, i = 1, . . . , d
for x
Γ j=1
which is divergence–free in Ω, and satisfies Green’s first formula

d
2µ eij (u, v)eij (v, x)dx = v(x) (T u)(x)dsx (6.65)
Ω i,j=1 Γ
for all v ∈ [H 1 (Ω)]d with div v = 0. The application of the interior trace
operator defines the single layer potential
int/ext
V := γ0 V : [H −1/2 (Γ )]d → [H 1/2 (Γ )]d
which allows a representation as given in (6.58). To investigate the ellipticity

of the single layer potential V we first note that u∗ = 0 and p = −1 defines
a solution of the homogeneous Stokes system. From the boundary integral
equation (6.62) we then obtain
1
(V t∗ )(x) = ( I + K)u∗ (x) = 0 for x ∈ Γ
2
with the associated boundary stress
⎛ ⎞d

d
t∗ (x) = −p∗ (x) n(x) + 2 ⎝ eij (u∗ , x)nj (x)⎠ = n(x) for x ∈ Γ.
j=1
i=1
Hence we can expect the ellipticity of the Stokes single layer potential V only
in a subspace which is orthogonal to the exterior normal vector n.
Let V L : H −1/2 (Γ ) → H 1/2 (Γ ) be the Laplace single layer potential which
is H −1/2 (Γ )–elliptic. Hence we can define

d
w, τ V L := V L wi , τi Γ
i=1
as an inner product in [H −1/2 (Γ )]d . When considering the subspace

! "
−1/2
[HV L (Γ )]d := w ∈ [H −1/2 (Γ )]d : w, nV L = 0
we can prove the following result [50, 159]:

−1/2
Theorem 6.39. The Stokes single layer potential V is [HV L (Γ )]d –elliptic,
i.e.
−1/2
V w, wΓ ≥ cV1 w2[H −1/2 (Γ )]d for all w ∈ [HV L (Γ )]d .
As for the homogeneous Neumann boundary value problem for the Laplace
equation we can introduce an extended bilinear form
V w, τ Γ := V w, τ Γ + w, nV L τ , nV L
which defines an [H −1/2 (Γ )]d –elliptic boundary integral operator V .

When the computational domain Ω is multiple connected, the dimension of
the kernel of the Stokes single layer potential is equal to the number of closed
sub–boundaries. Then we have to modify the stabilization in a corresponding
manner, see [116].

Finally we consider the interior Helmholtz equation
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω ⊂ Rd
where the fundamental solution is, see (5.20) for d = 2 and (5.18) for d = 3,
⎧
⎪
⎪
1
⎨ Y0 (k|x − y|) for d = 2,
∗ 2π
Uk (x, y) =
⎪
⎪ 1 eik|x−y|
⎩ for d = 3.
4π |x − y|
Then we can define the standard boundary integral operators for x ∈ Γ , i.e.
the single layer potential

(Vk w)(x) = Uk∗ (x, y)w(y)dsy ,
Γ
the double layer potential

∂
(Kk v)(x) = U ∗ (x, y)v(y)dsy ,
∂ny k
Γ
the adjoint double layer potential

∂
(Kk v)(x) = U ∗ (x, y)v(y)dsy ,
∂nx k
Γ
and the hypersingular boundary integral operator

∂ ∂
(Dk v)(x) = − U ∗ (x, y)v(y)dsy .
∂nx ∂ny k
Γ
As for the Laplace operator there hold all the mapping properties as given in
Theorem 6.34. In particular, Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded, but not
H −1/2 (Γ )–elliptic. However, the single layer potential is coercive satisfying a
Gårdings inequality.
For x ∈ Ω we consider the function

u(x) = (Vk w)(x) − (V w)(x) = [Uk∗ (x, y) − U ∗ (x, y)]w(y)dsy (6.66)
Ω
∗
where U (x, y) is the fundamental solution of the Laplace operator. In par-
ticular we have for y ∈ Γ and x ∈ Ω
−∆x Uk∗ (x, y) − k 2 Uk∗ (x, y) = 0, −∆x U ∗ (x, y) = 0.

Then, by interchanging differentiation and integration, we obtain

[−∆x − k 2 ]u(x) = [−∆x − k 2 ][Uk∗ (x, y) − U ∗ (x, y)]w(y)dsy
Γ

= k2 U ∗ (x, y)w(y)dsy .
Γ
Moreover,

−∆x [−∆x − k 2 ]u(x) = −k 2 ∆x U ∗ (x, y)w(y)dsy = 0,
Γ
i.e. the function u as defined in (6.66) solves the partial differential equation
−∆x [−∆x − k 2 ]u(x) = 0 for x ∈ Ω
which is of fourth order. Hence, we obtain as in the case of the Laplace oper-
ator, by considering the corresponding Newton potentials, that
Vk − V : H −1/2 (Γ ) → H 3 (Ω).
Thus,
Vk − V = γ0int [Vk − V ] : H −1/2 (Γ ) → H 5/2 (Γ ),
and by the compact imbedding of H 5/2 (Γ ) in H 1/2 (Γ ) we conclude that
Vk − V : H −1/2 (Γ ) → H 1/2 (Γ ) (6.67)
is compact.
Theorem 6.40. The single layer potential Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is

coercive, i.e. there exists a compact operator C : H −1/2 (Γ ) → H 1/2 (Γ ) such
that the Gårdings inequality
Vk w, wΓ + Cw, w, Γ ≥ cV1 w2H −1/2 (Γ ) for w ∈ H −1/2 (Γ )
is satisfied. For d = 2 we have to assume diam Ω < 1.
Proof. By considering the compact operator
C = V − Vk : H −1/2 (Γ ) → H 1/2 (Γ )
we have
Vk w, wΓ + Cw, wΓ = V w, wΓ ≥ cV1 w2H −1/2 (Γ )
by using Theorem 6.22 for d = 3 and Theorem 6.23 for d = 2.

Note that also the operators
6.10 Exercises 169
Dk − D : H 1/2 (Γ ) → H −1/2 (Γ ),
Kk − K : H 1/2 (Γ ) → H 1/2 (Γ ),

KK − K : H −1/2 (Γ ) → H −1/2 (Γ )
are compact where D, K, K are the hypersingular integral operator, the dou-
ble layer potential and its adjoint of the Laplace operator, respectively.
As for the Laplace operator the bilinear form of the hypersingular bound-
ary integral operator Dk can be written as, by using integration by parts
[107],
ı k|x−y|
1 e
Dk u, vΓ = curlΓ u(y), curlΓ v(x) dsy dsx
4π |x − y|
Γ Γ

k2 eı k|x−y|
− u(y)v(x) n(x), n(y) dsy dsx . (6.68)
4π |x − y|
Γ Γ
6.10 Exercises
Let Γ = ∂Ω be the boundary of the circle Ω = Br (0) ⊂ R2 which can be
described by using polar coordinates as
& '
cos 2πt
x(t) = r ∈ Γ for t ∈ [0, 1).
sin 2πt
6.1 Find a representation of the two–dimensional single layer potential

1
(V w)(x) = − log |x − y|w(y)dsy for x ∈ Γ = ∂Br (0)
2π
Γ
when using polar coordinates x = x(τ ) and y = y(t), respectively.

6.2 Find a representation of the two–dimensional double layer potential

1 (y − x, n(y))
(Kv)(x) = − v(y)dsy for x ∈ Γ = ∂Br (0)
2π |x − y|2
Γ
when using polar coordinates x = x(τ ) and y = y(t), respectively.

6.3 The eigenfunctions of the double layer potential as considered in Exercise
6.2 are given by
vk (t) = ei2πkt for k ∈ N0 .
Compute the associated eigenvalues.
6.4 By using the eigenfunctions as given in Exercise 6.3 compute all eigen-
values of the single layer potential as given in Exercise 6.1. Give sufficient
conditions such that the single layer potential is invertible, and positive
definite.
6.5 Prove Corollary 6.19.
6.6 Determine the eigenfunctions of the hypersingulur boundary integral op-
erator D for x ∈ ∂Br (0) and compute the corresponding eigenvalues.
6.7 Let now Γ be the boundary of an ellipse given by the parametrization
& '
a cos 2πt
x(t) = ∈ Γ for t ∈ [0, 2π).
b sin 2πt
Find a representation of the corresponding double layer potential

1 (y − x, n(y))
(Kv)(x) = − v(y)dsy for x ∈ Γ.
2π |x − y|2
Γ
6.8 Prove that the eigenfunctions of the double layer potential as considered
in Exercise 6.7 are given by
⎧
⎪
⎨ cos 2πkt for k > 0,
vk (t) = 1 for k = 0,
⎪
⎩
sin 2πkt for k < 0.
Compute the corresponding eigenvalues. Describe the behavior of the maximal

eigenvalue as ab → ∞.
6.9 Prove for the double layer potential of the Helmholtz equation that

Kk = K−k
is satisfied when considering the complex inner product

w, vΓ = w(x)v(x)dsx .
Γ
7
Boundary Integral Equations
In this chapter we consider boundary value problems for scalar homogeneous

partial differential equations
(Lu)(x) = 0 for x ∈ Ω (7.1)
where L is an elliptic and self–adjoint partial differential operator of sec-

ond order, and Ω is a bounded and simple connected domain with Lipschitz
boundary Γ = ∂Ω. In particular we focus on the Laplace and on the Helmholtz
equations. Note that boundary integral equations for boundary value prob-
lems in linear elasticity can be formulated and analyzed as for the Laplace
equation. To handle inhomogeneous partial differential equations, Newton po-
tentials have to be considered in addition. By computing particular solutions
of the inhomogeneous partial differential equations all Newton potentials can
be reduced to surface potentials only, see, for example, [86, 136].
Any solution u of the homogeneous partial differential equation (7.1) is
∈ Ω by the representation formula (5.2),
given for x

u(x) = U ∗ ( int U ∗ (
x, y)γ1int u(y)dsy − γ1,y x, y)γ0int u(y)dsy . (7.2)
Γ Γ
Hence we have to find the complete Cauchy data γ0int u(x) and γ1int u(x) for
x ∈ Γ , which are given by boundary conditions only partially. For this we
will describe appropriate boundary integral equations. The starting point is
the representation formula (7.2) and the related system (6.22) of boundary
integral equations,
& ' & '& '
γ0int u [1 − σ]I − K V γ0int u
= . (7.3)
γ1int u D σI + K γ1int u
This approach is called direct where the density functions of all boundary
integral operators are just the Cauchy data [γ0int u(x), γ1int u(x)], x ∈ Γ . When
172 7 Boundary Integral Equations
describing the solution of boundary value problems by using suitable poten-

tials, we end up with the so called indirect approach. For example, solutions
of the homogeneous partial differential equation (7.1) are given either by the
single layer potential

u(x) = U ∗ ( ∈ Ω,
x, y)w(y)dsy for x (7.4)
Γ
or by the double layer potential

u(x) = − γ1,y int U ∗ (
x, y)v(y)dsy ∈ Ω.
for x (7.5)
Γ
It is worth to mention that in general the density functions w and v of the

indirect approach have no physical meaning.
In this chapter we will consider different boundary integral equations to
find the unknown Cauchy data to describe the solution of several boundary
value problems with different boundary conditions.
7.1 Dirichlet Boundary Value Problem

First we consider the Dirichlet boundary value problem
(Lu)(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ. (7.6)
When using the direct approach (7.2) we obtain the representation formula

u(
x) = U ∗ ( int U ∗ (
x, y)γ1int u(y)dsy − γ1,y ∈ Ω (7.7)
x, y)g(y)dsy for x
Γ Γ
where we have to find the yet unknown Neumann datum γ1int u ∈ H −1/2 (Γ ).
By using the first boundary integral equation in (7.3) we obtain with
(V γ1int u)(x) = σ(x)g(x) + (Kg)(x) for x ∈ Γ (7.8)
a first kind Fredholm boundary integral equation. Since the single layer poten-
tial V : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded (see (6.8)) and H −1/2 (Γ )–elliptic
(see Theorem 6.22 for d = 3 and Theorem 6.23 for d = 2 when assuming
diam(Ω) < 1), we conclude the unique solvability of the boundary integral
equation (7.8) when applying the Lax–Milgram lemma (Theorem 3.4). More-
over, the unique solution γ1int u ∈ H −1/2 (Γ ) satisfies
1 cW
2
γ1int uH −1/2 (Γ ) ≤ (σI + K)gH 1/2 (Γ ) ≤ gH 1/2 (Γ ) .
cV1 cV1
7.1 Dirichlet Boundary Value Problem 173
Since the boundary integral equation (7.8) is formulated in H 1/2 (Γ ), this gives
0 = V γ1int u − (σI + K)gH 1/2 (Γ )

V γ1int u − (σI + K)g, τ Γ
= sup .
0=τ ∈H −1/2 (Γ ) τ H −1/2 (Γ )
and therefore, instead of (7.8) we may consider the equivalent variational

problem to find γ1int u ∈ H −1/2 (Γ ) such that
1
V γ1int u, τ Γ = ( I + K)g, τ Γ (7.9)
2
is satisfied for all τ ∈ H −1/2 (Γ ). Note that the definition of σ(x) gives σ(x) =
1
2 for almost all x ∈ Γ .
Instead of (7.8) we may also use the second boundary integral equation in
(7.3) to find the unknown Neumann datum γ1int u ∈ H −1/2 (Γ ), i.e.
([1 − σ]I − K )γ1int u(x) = (Dg)(x) for x ∈ Γ (7.10)
which is a second kind Fredholm boundary integral equation. The solution of

this boundary integral equation is given by the Neumann series
∞

γ1int u(x) = (σI + K ) (Dg)(x) for x ∈ Γ. (7.11)
=0
The convergence of the series (7.11) in H −1/2 (Γ ) follows from the contraction
property (6.54) of σI +K when considering the equivalent Sobolev norm ·V
which is induced by the single layer potential V .
When using the indirect single layer potential ansatz (7.4) to find the un-
known density w ∈ H −1/2 (Γ ) we have to solve the boundary integral equation
(V w)(x) = g(x) for x ∈ Γ. (7.12)
Note that the boundary integral equation (7.12) differs from the boundary
integral equation (7.8) of the direct approach only in the definition of the
right hand side. Hence we can conclude the unique solvability of the boundary
integral equation (7.12) as for (7.8).
By using the double layer potential (7.5) to describe the solution of the
homogeneous partial differential equation we obtain from the jump relation
(6.14) of the double layer potential the boundary integral equation
[1 − σ(x)]v(x) − (Kv)(x) = g(x) for x ∈ Γ (7.13)
to compute the density v ∈ H 1/2 (Γ ) via the Neumann series

∞

v(x) = (σI + K) g(x) for x ∈ Γ. (7.14)
=0
The convergence of the series (7.14) in H 1/2 (Γ ) follows from the contraction
property (6.51) of σI + K when considering the equivalent Sobolev norm
· V −1 which is induced by the inverse single layer potential V −1 .
To obtain a variational formulation of the boundary integral equation
(7.13) in H 1/2 (Γ ) we first consider
12 v − Kv − g, τ Γ
0 = [1 − σ]v − Kv − gH 1/2 (Γ ) = sup
0=τ ∈H −1/2 (Γ ) τ H −1/2 (Γ )
where we have used σ(x) = 12 for almost all x ∈ Γ .

This gives a variational formulation to find v ∈ H 1/2 (Γ ) such that
1
( I − K)v, τ Γ = g, τ Γ (7.15)
2
is satisfied for all τ ∈ H −1/2 (Γ ).
Lemma 7.1. There holds the stability condition
( 12 I − K)v, τ Γ
cS vH 1/2 (Γ ) ≤ sup for all v ∈ H 1/2 (Γ )
0=τ ∈H −1/2 (Γ ) τ H −1/2 (Γ )
with a positive constant cS > 0.
Proof. Let v ∈ H 1/2 (Γ ) be arbitrary but fixed. For τv := V −1 v ∈ H −1/2 (Γ )

we then have
1
τv H −1/2 (Γ ) = V −1 vH −1/2 (Γ ) ≤ vH 1/2 (Γ ) .
cV1
By using the contraction estimate (6.51) and the mapping properties of the
single layer potential V we obtain
1 1
( I − K)v, τv Γ = ( I − K)v, V −1 vΓ
2 2
1
= V −1 v, vΓ − V −1 ( I + K)v, v, Γ
2
1
≥ v2V −1 − ( I + K)vV −1 vV −1
2
≥ (1 − cK ) v2V −1
= (1 − cK ) V −1 v, vΓ
1
≥ (1 − cK ) v2H 1/2 (Γ )
cV2
cV1
≥ (1 − cK ) vH 1/2 (Γ ) τv H −1/2 (Γ )
cV2
7.2 Neumann Boundary Value Problem 175
from which the stability condition follows immediately.

Hence we conclude the unique solvability of the variational problem (7.15)
by applying Theorem 3.7.
Remark 7.2. To describe the solution of the Dirichlet boundary value problem
(7.6) we have described four different boundary integral equations, and we
have shown their unique solvability. Depending on the application and on the
discretization scheme to be used, each of the above formulations may have
their advantages or disadvantages. In this book, we will mainly consider the
approximate solution of the variational formulation (7.9).
7.2 Neumann Boundary Value Problem

When considering the scalar Neumann boundary value problem
(Lu)(x) = 0 for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ (7.16)
we have to assume the solvability condition (1.17),

g(x)dsx = 0 . (7.17)
Γ
The representation formula (7.2) then yields

∗
u(
x) = U (x, y)g(y)dsy − γ1int U ∗ (
x, y)γ0int u(y)dsy ∈ Ω (7.18)
for x
Γ Γ
where we have to find the yet unknown Dirichlet datum γ0int u ∈ H 1/2 (Γ ).
From the second boundary integral equation of the Calderon system (7.3) we
obtain
(Dγ0int u)(x) = (1 − σ(x))g(x) − (K g)(x) for x ∈ Γ (7.19)
which is a first kind Fredholm boundary integral equation. Due to (6.17) we
have that u0 ≡ 1 is an eigensolution of the hypersingular boundary integral
operator, i.e. (Du0 )(x) = 0. Hence we have ker D = span {u0 }, and to ensure
the solvability of the boundary integral equation (7.19) we need to assume,
by applying Theorem 3.6, the solvability condition
(1 − σ)g − K g ∈ Im(D) = (ker D)0 .
Note that (ker D)0 is the orthogonal space which is induced by ker D, see
(3.15). From
[1 − σ]g − K g, u0 Γ = g, 1Γ − (σI + K )g, u0 Γ (7.20)

= g, 1Γ − g, (σI + K)u0 Γ = 0
we then conclude the solvability of the boundary integral equation (7.19).

The hypersingular boundary integral operator D : H 1/2 (Γ ) → H −1/2 (Γ ) is
1/2
bounded (see (6.16)) and H∗ (Γ )–elliptic (see Theorem 6.24). Then, apply-
ing the Lax–Milgram lemma (Theorem 3.4), there exists a unique solution
1/2
γ0int u ∈ H∗ (Γ ) of the hypersingular boundary integral equation (7.19). The
1/2
equivalent varitional problem is to find γ0int u ∈ H∗ (Γ ) such that
1
Dγ0int u, vΓ = ( I − K )g, vΓ (7.21)
2
1/2
is satisfied for all v ∈ H∗ (Γ ).
Instead of the variational problem (7.21) with a constraint we may also
consider a saddle point problem to find (γ0int u, λ) ∈ H 1/2 (Γ ) × R such that
Dγ0int u, vΓ + λ v, weq Γ = ( 12 I − K )g, vΓ
(7.22)
γ0int u, weq Γ =0
is satisfied for all v ∈ H 1/2 (Γ ).
When inserting v = u0 ∈ H 1/2 (Γ ) as a test function of the first equation in
the saddle point problem (7.22) this gives Du0 = 0 and from the orthogonality
(7.20) we get
0 = λ 1, weq Γ = λ 1, V −1 1Γ
and therefore λ = 0, since the inverse single layer potential V −1 is elliptic.
The saddle point problem (7.22) is therefore equivalent to finding (γ0int u, λ) ∈
H 1/2 (Γ ) × R such that
Dγ0int u, vΓ + λ v, weq Γ = ( 12 I − K )g, vΓ
(7.23)
γ0int u, weq Γ − λ/α =0
is satisfied for all v ∈ H 1/2 (Γ ). Here, α ∈ R+ is some parameter to be chosen.
Hence we can eliminate the Lagrange multiplier λ ∈ R to obtain a modified
variational problem to find γ0int u ∈ H 1/2 (Γ ) such that
1
Dγ0int u, vΓ + α γ0int u, weq Γ v, weq Γ = ( I − K )g, vΓ (7.24)
2
is satisfied for all v ∈ H 1/2 (Γ ). The modified hypersingular boundary integral
operator D : H 1/2 (Γ ) → H −1/2 (Γ ) which is defined via the bilinear form
vΓ := Dw, vΓ + α w, weq Γ v, weq Γ
Dw,
for all v, w ∈ H 1/2 (Γ ) is bounded, and H 1/2 (Γ )–elliptic, due to
vΓ = Dv, vΓ + α [v, weq Γ ]2
Dv,
2 2
≥ c̄D
1 |v|H 1/2 (Γ ) + α [v, weq Γ ]
! "
2 2
≥ min{c̄D1 , α} |v|H 1/2 (Γ ) + [v, weq Γ ]
2 2
1 , α} vH 1/2 (Γ ) ≥ ĉ1 vH 1/2 (Γ )
= min{c̄D D
∗
for all v ∈ H 1/2 (Γ ). This estimate also indicates an appropriate choice of

the parameter α ∈ R+ . Note that the modified variational problem (7.24)
admits a unique solution for any right hand side, and therefore for any given
Neumann datum g ∈ H −1/2 (Γ ). If the given Neumann datum g satisfies the
solvability condition (7.17), then we conclude, by inserting v = u0 ≡ 1 as a
test function, from the variational problem (7.24)
α γ0int u, weq Γ 1, weq Γ = 0, 1, weq Γ = 1, V −1 1Γ > 0

1/2
and therefore γ0int u ∈ H∗ (Γ ). The modified variational problem (7.24) is
thus equivalent to the original variational problem (7.21).
1/2
Since the hypersingular boundary integral operator D is also H∗∗ (Γ )–
1/2
elliptic (see (6.38)), there also exists a unique solution γ0int u ∈ H∗∗ (Γ ) of
the boundary integral equation (7.19). In analogy to the above treatment we
obtain a modified variational problem to find γ0int u ∈ H 1/2 (Γ ) such that
1
Dγ0int u, vΓ + ᾱγ0int u, 1Γ v, 1Γ = ( I − K )g, vΓ (7.25)
2
is satisfied for all v ∈ H 1/2 (Γ ). Again, ᾱ ∈ R+ is some parameter to be chosen.
Moreover, if we assume the solvability condition (7.17) this gives γ0int u ∈
1/2
H∗∗ (Γ ). By the bilinear form
D̂w, vΓ := Dw, vΓ + ᾱ w, 1Γ v, 1Γ (7.26)
for all w, v ∈ H 1/2 (Γ ) we define a modified hypersingular boundary integral

operator D̂ : H 1/2 (Γ ) → H −1/2 (Γ ) which is bounded and H 1/2 (Γ )–elliptic.
When using the indirect double layer potential (7.5) to find the unknown
1/2
density v ∈ H∗ (Γ ) we obtain the hypersingular boundary integral equation
(Dv)(x) = g(x) for x ∈ Γ, (7.27)
which can be analyzed as the boundary integral equation (7.19).

If we consider the representation formula (7.18) of the direct approach, and
use the first boundary integral equation of the resulting Calderon projection
(7.3), we find the yet unknown Dirichlet datum as the solution of the boundary
integral equation
(σI + K)γ0int u(x) = (V g)(x) for x ∈ Γ. (7.28)
The solution of the boundary integral equation (7.28) is given by the Neumann
series
∞

γ0int u(x) = ([1 − σ]I − K) (V g)(x) for x ∈ Γ. (7.29)
=0
The convergence of the Neumann series (7.29) follows from the contraction
1/2
property (6.52) of ([1 − σ]I − K) in H∗ (Γ ) when considering the equivalent
Sobolev norm · V −1 which is induced by the inverse single layer potential

V −1 . The variational formulation of the boundary integral equation (7.28)
needs therefore to be considered in H 1/2 (Γ ). Since the single layer potential
V : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded and H −1/2 (Γ )–elliptic, we can define
w, vV −1 := V −1 w, vΓ for w, v ∈ H 1/2 (Γ )
to be an inner product in H 1/2 (Γ ). The variational formulation of the bound-

ary integral equation (7.28) with respect to the inner product ·, ·V −1 then
1/2
reads to find γ0int u ∈ H∗ (Γ ) such that
(σI + K)γ0int u, vV −1 = V g, vV −1 (7.30)

1/2
is satisfied for all v ∈ H∗ (Γ ). The variational problem (7.30) is equivalent
1/2
to finding γ0int u ∈ H∗ (Γ ) such that
Sγ0int u, vΓ = V −1 (σI + K)γ0int u, vΓ = g, vΓ (7.31)

1/2
is satisfied for all v ∈ H∗ (Γ ). Since the Steklov–Poincaré operator S :
H 1/2 (Γ ) → H −1/2 (Γ ) admits the same mapping properties as the hyper-
singular boundary integral operator D : H 1/2 (Γ ) → H −1/2 (Γ ), the unique
solvability of the variational problem (7.31) follows as for the variational
formulation (7.21). When using the symmetric representation (6.42) of the
Steklov–Poincaré operator S, the variational problem (7.31) is equivalent to
0 1
Sγ0int u, vΓ = D + (σI + K )V −1 (σI + K) γ0int u, vΓ = g, vΓ (7.32)
When using the indirect single layer potential (7.4) we finally obtain the
boundary integral equation
(σI + K )w(x) = g(x) for x ∈ Γ (7.33)
to find the unknown density w ∈ H −1/2 (Γ ) which is given via the Neumann
series
∞
w(x) = ((1 − σ)I − K ) g(x) for x ∈ Γ. (7.34)
=0
The convergence of the Neumann series (7.34) follows from the contraction
−1/2
property (6.55) of ((1−σ)I −K ) in H∗ (Γ ) when considering the equivalent
Sobolev norm · V .
Remark 7.3. For the solution of the Neumann boundary value problem (7.16)
again we have given and analyzed four different formulations of boundary
integral equations. As for the Dirichlet boundary value problem each of them
may have their advantages and disadvantages. Here we will mainly consider
the approximate solution of the modified variational formulation (7.25).
7.3 Mixed Boundary Conditions 179
7.3 Mixed Boundary Conditions

In addition to the standard boundary value problems (7.6) and (7.16) with
either Dirichlet or Neumann boundary conditions, boundary value problems
with mixed boundary conditions are of special interest,
(Lu)(x) = 0 for x ∈ Ω,
γ0int u(x) = gD (x) for x ∈ ΓD , (7.35)
γ1int u(x) = gN (x) for x ∈ ΓN .
∈Ω
From the representation formula (7.2) we get for x

u(
x) = U ∗ (
x, y)gN (y)dsy + U ∗ ( int u(y)ds
x, y)γ1,y y (7.36)
ΓN ΓD

− int U ∗ (
γ1,y x, y)gD (y)dsy − γ1int U ∗ (
x, y)γ0int u(y)dsy .
ΓD ΓN
Hence we have to find the yet unknown Dirichlet datum γ0int u(x) for x ∈ ΓN
and the Neumann datum γ1int u(x) for x ∈ ΓD . Keeping in mind the differ-
ent boundary integral formulations for both the Dirichlet and the Neumann
problems, there seems to be a wide variety of different boundary integral for-
mulations to solve the mixed boundary value problem (7.35). Here we will
only consider two formulations which are based on the representation formula
(7.36) of the direct approach.
The symmetric formulation [134] is based on the use of the first boundary
integral equation in (7.3) for x ∈ ΓD while the second boundary integral
equation in (7.3) is considered for x ∈ ΓN ,
(V γ1int u)(x) = (σI + K)γ0int u(x) for x ∈ ΓD ,

(7.37)
(Dγ0int u)(x) = ((1 − σ)I − K )γ1int u(x) for x ∈ ΓN .
Let gD ∈ H 1/2 (Γ ) and gN ∈ H −1/2 (Γ ) be suitable extensions of the given
boundary data gD ∈ H 1/2 (ΓD ) and gN ∈ H −1/2 (ΓN ) satisfying
gD (x) = gD (x) for x ∈ ΓD , gN (x) = gN (x) for x ∈ ΓN .
Inserting these extensions into the system (7.37) this gives the symmetric
formulation to find
1/2 (ΓN ),
:= γ0int u − gD ∈ H
u −1/2 (ΓD )
t := γ1int u − gN ∈ H
such that
(V
t)(x) − (K u gD (x) − (V gN )(x)
)(x) = (σI + K) for x ∈ ΓD ,
u)(x) + (K
(D t)(x) = ((1 − σ)I − K )
gN (x) − (D
gD )(x) for x ∈ ΓN .
(7.38)
The related variational formulation is to find ( −1/2 (ΓD ) × H

) ∈ H
t, u 1/2 (ΓN )
such that
a( ; τ, v) = F (τ, v)
t, u (7.39)
−1/2 (ΓD ) × H
is satisfied for all (τ, v) ∈ H 1/2 (ΓN ) where
a( ; τ, v) = V
t, u , τ ΓD + K
t, τ ΓD − K u t, vΓN + D
u, vΓN ,
1 1
gD − V gN , τ ΓD + ( I − K )
F (τ, v) = ( I + K) gN − D
gD , vΓN .
2 2
Lemma 7.4. The bilinear form a(·; ·) of the symmetric boundary integral for-
−1/2 (ΓD ) × H
mulation is bounded and H 1/2 (ΓN )–elliptic, i.e.
a(t, u; τ, v) ≤ cA
2 (t, u)H -1/2 (ΓN ) (τ, v)H
-−1/2 (ΓD )×H -−1/2 (ΓD )×H
-1/2 (ΓN )
and
2
1 } (τ, v)H
a(τ, v; τ, v) ≥ min{cV1 , ĉD -−1/2 (Γ -1/2 (ΓN )
D )×H
−1/2 (ΓD ) × H
for all (t, u), (τ, v) ∈ H 1/2 (ΓN ) where the norm is defined by
(τ, v)2H
-−1/2 (Γ -1/2 (ΓN ) := τ 2H
-−1/2 (Γ + v2H
-1/2 (Γ .
D )×H D) N)
Proof. By using
a(τ, v; τ, v) = V τ, τ ΓD − Kv, τ ΓD + K τ, vΓN + Dv, vΓN

= V τ, τ ΓD + Dv, vΓN
≥ cV1 τ 2H
-−1/2 (Γ + ĉD 2
1 vH
-1/2 (Γ
D) N)
we conclude the ellipticity of the bilinear form a(·, ·; ·, ·) from the ellipticity
estimates of the boundary integral operators V and D, see Theorem 6.22 for
d = 3 and Theorem 6.23 for d = 2, as well as (6.39). The boundedness of
the bilinear form a(·, ·; ·, ·) is a direct consequence of the boundedness of all
boundary integral operators.
Since the linear form F (τ, v) is bounded for (τ, v) ∈ H −1/2 (ΓD ) ×
H (ΓN ), the unique solvability of the variational formulation (7.39) follows
1/2
from the Lax–Milgram lemma (Theorem 3.4).

To obtain a second boundary integral equation to solve the mixed bound-
ary value problem (7.35) we consider the Dirichlet to Neumann map (6.44) to
find γ0int u ∈ H 1/2 (Γ ) such that

γ1int u(x) = (Sγ0int u)(x) = gN (x) for x ∈ ΓN .
Let gD ∈ H 1/2 (Γ ) be some arbitrary but fixed extension of the given Dirichlet
datum gD ∈ H 1/2 (ΓD ). Then we have to find u 1/2 (ΓN )
:= γ0int u − gD ∈ H
such that
7.5 Exterior Boundary Value Problems 181
S
u, vΓN = gN − S
gD , vΓN (7.40)
1/2
is satisfied for all v ∈ H (ΓN ). Since the Steklov–Poincaré operator S :
1/2 (ΓN )–elliptic (see (6.48)) we con-
H 1/2 (Γ ) → H −1/2 (Γ ) is bounded and H
clude the unique solvability of the variational problem (7.40) from the Lax–
Milgram lemma (Theorem 3.4). If the Dirichlet datum γ0int u ∈ H 1/2 (Γ ) is
known, we can compute the complete Neumann datum γ1int u ∈ H −1/2 (Γ ) by
solving a Dirichlet boundary value problem.
7.4 Robin Boundary Conditions

Next we consider the Robin boundary value problem
(Lu)(x) = 0 for x ∈ Ω, γ1int u(x) + κ(x)γ0int u(x) = g(x) for x ∈ Γ.
To formulate a boundary integral equation to find the yet unknown Dirichlet

datum γ0int u ∈ H 1/2 (Γ ) again we can use the Dirichlet to Neumann map
(6.44), i.e.
γ1int u(x) = (Sγ0int u)(x) = g(x) − κ(x)γ0int u(x) for x ∈ Γ.
The related variational problem is to find γ0int u ∈ H 1/2 (Γ ) such that
Sγ0int u, vΓ + κγ0int u, vΓ = g, vΓ (7.41)
is satisfied for all v ∈ H 1/2 (Γ ). By using (6.45) and the H 1/2 (Γ )–semi–
ellipticity of the hypersingular boundary integral operator D and assuming
κ(x) ≥ κ0 for x ∈ Γ we conclude
a(v, v) := Sv, vΓ + κv, vΓ

2 2 2
≥ c̄D
1 |v|H 1/2 (Γ ) + κ0 vΓ = min{c̄1 , κ0 } vH 1/2 (Γ )
D
and therefore the H 1/2 (Γ )–ellipticity of the bilinear form a(·, ·). Again we
obtain the unique solvability of the variational problem (7.41) from the Lax–
Milgram lemma (Theorem 3.4).
7.5 Exterior Boundary Value Problems

An advantage of boundary integral equation methods is the explicit consider-
ation of far field boundary conditions when solving boundary value problems
in the exterior domain Ω c := Rd \Ω. As a model problem we consider the
exterior Dirichlet boundary value problem for the Laplace equation,
−∆u(x) = 0 for x ∈ Ω c , γ0ext u(x) = g(x) for x ∈ Γ (7.42)

together with the far field boundary condition

1
|u(x) − u0 | = O as |x| → ∞. (7.43)
|x|
where u0 ∈ R is some given number.
First we consider Green’s first formula for the exterior domain. For y0 ∈ Ω
and R ≥ 2 diam(Ω) let BR (y0 ) be a ball with center y0 , which circumscribes
Ω. Using the representation formula (6.1) for x ∈ BR (y0 )\Ω this gives

∗ ext ext U ∗ (x, y)γ ext u(y)ds
u(x) = − U (x, y)γ1 u(y)dsy + γ1,y 0 y
Γ Γ

∗ int U ∗ (x, y)γ int u(y)ds .
+ U (x, y)γ1int u(y)dsy − γ1,y 0 y
∂BR (y0 ) ∂BR (y0 )
Inserting the far field boundary condition (7.43) and taking the limit R → ∞
this results in the representation formula for x ∈ Ω c ,

u(x) = u0 − U ∗ (x, y)γ1ext u(y)dsy + γ1,y ext U ∗ (x, y)γ ext u(y)ds .
0 y
Γ Γ
To find the unknown Cauchy data again we can formulate different boundary
integral equations. The application of the exterior trace operator gives
γ0ext u(x) = u0 − (V γ1ext u)(x) + σ(x)γ0ext u(x) + (Kγ0ext u)(x) for x ∈ Γ,
while the application of the exterior conormal derivative yields
γ1ext u(x) = [1 − σ(x)]γ1ext u(x) − (K γ1ext u)(x) − (Dγ0ext u)(x) for x ∈ Γ.
As in (6.22) we obtain a system of boundary integral equations,

& ' & '& ' & '
γ0ext u σI + K −V γ0ext u u0
=
+ .
ext
γ1 u −D [1 − σ]I − K ext
γ1 u 0
Using the boundary integral equations of this system, the exterior Calderon
projection, we can formulate different boundary integral equations to handle
exterior boundary value problems with different boundary conditions. In par-
ticular for the exterior Dirichlet boundary value problem (7.42) and (7.43) we
can find the yet unknown Neumann datum γ1ext u ∈ H −1/2 (Γ ) as the unique
solution of the boundary integral equation
(V γ1ext u)(x) = −[1 − σ(x)]gD (x) + (KgD )(x) + u0 for x ∈ Γ. (7.44)
Note that the unique solvability of the boundary integral equation (7.44)
follows as for interior boundary value problems from the mapping properties
of the single layer potential V : H −1/2 (Γ ) → H 1/2 (Γ ).

Finally we consider boundary value problems for the Helmholtz equation, i.e.
the interior Dirichlet boundary value problem
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ (7.45)
where the solution is given by the representation formula

∗ int int U ∗ (x, y)g(y)ds
u(x) = Uk (x, y)γ1 u(y)dsy − γ1,y k y for x ∈ Ω.
Γ Γ
The unknown Neumann datum t = γ1int u ∈ H −1/2 (Γ ) solves the boundary

integral equation
1
(Vk t)(x) = ( I + Kk )g(x) for x ∈ Γ. (7.46)
2
Since the single layer potential Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is coercive, see
Theorem 6.40, we can apply Theorem 3.15 to investigate the solvability of the
boundary integral equation (7.46).
Lemma 7.5. If k 2 = λ is an eigenvalue of the Dirichlet eigenvalue problem

of the Laplace equation,
−∆uλ (x) = λ uλ (x) for x ∈ Ω, γ0int uλ (x) = 0 for x ∈ Γ, (7.47)
then the single layer potential Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is not injective, i.e.
(Vk γ1int uλ )(x) = 0 for x ∈ Γ.
Moreover,
1
( I − Kk )γ1int uλ (x) = 0 for x ∈ Γ.
2
Proof. The assertion immediately follows from the direct boundary integral
equations
1
(Vk γiint uλ )(x) = ( I + Kk )γ0int uλ (x) = 0 for x ∈ Γ
2
and
1
( I − Kk )γiint uλ (x) = (Dk γ0int uλ )(x) = 0 .

2
Hence we conclude, that if k 2 is not an eigenvalue of the Dirichlet eigen-

value problem of the Laplace equation, the Helmholtz single layer potential
Vk : H −1/2 (Γ ) → H 1/2 (Γ ) is coercive and injective, i.e. the boundary integral
equation (7.46) admits a unique solution, see Theorem 3.15.
Next we consider the exterior Dirichlet boundary value problem
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω c , γ0ext u(x) = g(x) for x ∈ Γ (7.48)
where, in addition, we have to require the Sommerfeld radiation condition

x 1

|x| , ∇u(x) − iku(x) = O |x|2 as |x| → ∞
Note that the exterior Dirichlet boundary value problem is uniquely solv-
able due to the Sommerfeld radiation condition. The solution is given by the
representation formula

u(x) = − Uk∗ (x, y)γ1ext u(y)dsy + γ1,yext U ∗ (x, y)g(y)ds
k y for x ∈ Ω c .
Γ Γ
To find the unknown Neumann datum t = γ1ext u ∈ H −1/2 (Γ ) we consider the

direct boundary integral equation
1
(Vk t)(x) = (− I + Kk )g(x) for x ∈ Γ. (7.49)
2
Since the single layer potential Vk of the exterior Dirichlet boundary value
problem coincides with the single layer potential of the interior problem, Vk is
not invertible when k 2 = λ is an eigenvalue of the Dirichlet eigenvalue problem
(7.47). However, due to
1 1
(− g, γ1int )uλ Γ = −g, ( I − K−k )γ1int uλ Γ = 0
2 2
we conclude
1
(− I + Kk )g ∈ Im Vk .
2
In fact, the boundary integral equation (7.49) of the direct approach is solvable
but not unique.
Instead of a direct approach, we may also consider an indirect single layer
potential approach

u(x) = Uk∗ (x, y)w(y)dsy for x ∈ Ω c
Γ
which leads to the boundary integral equation to find w ∈ H −1/2 (Γ ) such

that
(Vk w)(x) = g(x) for x ∈ Γ. (7.50)
As before, we have unique solvability of the boundary integral equation (7.50)
only for those wave numbers k 2 which are not eigenvalues of the interior
Dirichlet eigenvalue problem (7.47).
When using an indirect double layer potential ansatz

u(x) = ext U ∗ (x, y)v(y)ds
γ1,y for x ∈ Ω c
k y
Γ
the unknown density function v ∈ H 1/2 (Γ ) solves the boundary integral

equation
1
( I + Kk )v(x) = g(x) for x ∈ Γ. (7.51)
2
Lemma 7.6. If k 2 = µ is an eigenvalue of the interior Neumann eigenvalue
problem of the Laplace equation,
−∆uµ (x) = µuµ (x) for x ∈ Ω, γ1int uµ (x) = 0 for x ∈ Γ, (7.52)
then
1
( I + Kk )γ0int uµ (x) = 0 for x ∈ Γ.
2
Proof. The assertion immediately follows from the direct boundary integral
equation
1
( I + Kk )γ0int uµ (x) = (Vk γ1int uµ )(x) = 0 for x ∈ Γ.

2
The boundary integral equation (7.51) is therefore uniquely solvable if k 2 is

not an eigenvalue of the Neumann eigenvalue problem (7.52).
Although the exterior Dirichlet boundary value problem for the Helmholtz
equation is uniquely solvable, the related boundary integral equations may
not be solvable, in particular, when k 2 is either an eigenvalue of the interior
Dirichlet eigenvalue problem (7.47), or of the interior Neumann eigenvalue
problem (7.52). Since k 2 can not be an eigenvalue of both the interior Dirichlet
and the interior Neumann boundary value problem, one may combine both the
indirect single and double layer potential formulations to derive a boundary
integral equation which is uniquely solvable for all wave numbers. This leads
to the well known Brakhage–Werner formulation [23]

u(x) = ext U ∗ (x, y)w(y)ds − iη
γ1,y Uk∗ (x, y)w(y)dsy for x ∈ Ω c
k y
Γ Γ
where η ∈ R+ is some real parameter. This leads to a boundary integral

equation to find w ∈ L2 (Γ ) such that
1
( I + Kk )w(x) − iη(Vk w)(x) = g(x) for x ∈ Γ. (7.53)
2
Instead of considering the boundary integral equation (7.53) in L2 (Γ ), one
may formulate some modified boundary integral equations to be considered
in the energy space H −1/2 (Γ ), i.e. find w ∈ H −1/2 (Γ ) such that

1 −1 1
Vk + iη I + Kk D I + K−k w(x) = g(x) for x ∈ Γ (7.54)
2 2
where D is the modified hypersingular integral operator of the Laplace equa-

tion as defined in (7.24). The modified boundary integral equation (7.54)
admits a unique solution for all wave number k for general Lipschitz domains
[54], for other regularizations, see [33, 34].
7.7 Exercises
7.1 Consider the mixed boundary value problem
−∆u(x) = 0 for x ∈ Ω\Ω0 , Ω0 ⊂ Ω,

γ0int u(x) = g(x) for x ∈ Γ = ∂Ω,
γ int u(x) = 0
1 for x ∈ Γ0 = ∂Ω0 .
Derive the symmetric formulation of boundary integral equations to find the

complete Cauchy data. Discuss the solvability of the resulting variational
problem.
7.2 Discuss boundary integral formulations to solve the exterior Neumann
−∆u(x) − k 2 u(x) = 0 for x ∈ Ω c , γ1ext u(x) = g(x) for x ∈ Γ
and
x 1

|x| , ∇u(x) − iku(x) = O |x|2 as |x| → ∞.
8
Approximation Methods
In this chapter we describe and analyze approximation methods to solve the

variational problems for operator equations as formulated in Chapter 3. This
is done by introducing conforming finite dimensional trial spaces leading to
linear systems of algebraic equations.
8.1 Galerkin–Bubnov Methods

Let A : X → X be a bounded and X–elliptic linear operator satisfying
2
Av, v ≥ cA
1 vX , AvX ≤ cA
For a given f ∈ X we want to find the solution u ∈ X of the variational

problem (3.4),
Au, v = f, v for all v ∈ X. (8.1)
Due to the Lax–Milgram theorem 3.4 there exists a unique solution of the
variational problem (8.1) satisfying
1
uX ≤ f X .
cA
1
For M ∈ N we consider a sequence
XM := span{ϕk }M
k=1 ⊂ X
of conforming trial spaces. The approximate solution

M
uM := u k ϕk ∈ X M (8.2)
k=1
is defined as the solution of the Galerkin–Bubnov variational problem

188 8 Approximation Methods
AuM , vM = f, vM for all vM ∈ XM . (8.3)
Note that we have used the same trial and test functions for the Galerkin–
Bubnov method.
It remains to investigate the unique solvability of the variational problem
(8.3), the stability of the approximate solutions uM ∈ XM as well as their
convergence for M → ∞ to the unique solution u ∈ X of the variational
problem (8.1). Due to XM ⊂ X we can choose v = vM ∈ XM in the variational
formulation (8.1). Subtracting the Galerkin–Bubnov problem (8.3) from the
continuous variational formulation (8.1) this gives the Galerkin orthogonality
A(u − uM ), vM = 0 for all vM ∈ XM . (8.4)
Inserting the approximate solution (8.2) into the Galerkin–Bubnov formu-

lation (8.3) we obtain, by using the linearity of the operator A, the finite
dimensional variational problem

M
uk Aϕk , ϕ = f, ϕ for = 1, . . . , M.
k=1
With
AM [, k] := Aϕk , ϕ , f := f, ϕ
for k, = 1, . . . , M this is equivalent to the linear system of algebraic equations
AM u = f (8.5)
to find the coefficient vector u ∈ RM . For any arbitrary vector v ∈ RM we

can define a function

M
vM = v k ϕk ∈ X M
k=1
and vice versa. For arbitrary given vectors u, v ∈ RM we then have

M
M
M
M
(AM u, v) = AM [, k]uk v = Aϕk , ϕ uk v
k=1 =1 k=1 =1

M
M
= A u k ϕk , v ϕ = AuM , vM .
k=1 =1
Hence, all properties of the operator A : X → X are inherited by the stiffness

matrix AM ∈ RM ×M . In particular, the matrix AM is symmetric and posi-
tive definite, since the operator A is self–adjoint and X–elliptic, respectively.
Indeed,
2
(AM v, v) = AvM , vM ≥ cA 1 vM X
8.1 Galerkin–Bubnov Methods 189
for all v ∈ RM ↔ vM ∈ XM implies that AM is positive definite. Therefore,

the X–ellipticity of the operator A implies the unique solvability of the varia-
tional problem (8.1) as well as the unique solvability of the Galerkin–Bubnov
formulation (8.3) and hence, of the equivalent linear system (8.5).
Theorem 8.1 (Cea’s Lemma). Let A : X → X be a bounded and X–

elliptic linear operator. For the unique solution uM ∈ XM of the variational
problem (8.3) there holds the stability estimate
1
uM X ≤ f X (8.6)
cA
1
as well as the error estimate

cA
2
u − uM X ≤ inf u − vM X . (8.7)
cA
1 vM ∈XM
Proof. The unique solvability of the variational problem (8.3) was already
discussed before. For the approximate solution uM ∈ XM of (8.3) we conclude
from the X–ellipticity of A
2
1 uM X ≤ AuM , uM = f, uM ≤ f X uM X
cA
and therefore we obtain the stability estimate (8.6). By using the X–ellipticity
and the boundedness of the linear operator A, and by using the Galerkin
orthogonality (8.4) we get for any arbitrary vM ∈ XM
2
1 u − uM X ≤ A(u − uM ), u − uM
cA
= A(u − uM ), u − vM + A(u − uM ), vM − uM
= A(u − uM ), u − vM
≤ cA
2 u − uM X u − vM X
and therefore the error estimate (8.7).

The convergence of the approximate solution uM → u ∈ X as M → ∞
then follows from the approximation property of the trial space XM ,
lim inf v − vM X = 0 for all v ∈ X. (8.8)

M →∞ vM ∈XM
The sequence of conforming trial spaces {XM }M ∈N ⊂ X has to be constructed

in such a way that the approximation property (8.8) can be ensured. In Chap-
ter 9 we will consider the construction of local polynomial basis functions for
finite elements, while in Chapter 10 we will do the same for boundary el-
ements. Assuming additional regularity of the solution, we then prove also
corresponding approximation properties.
8.2 Approximation of the Linear Form

In different applications the right hand side f ∈ X is given as f = Bg where
g ∈ Y is prescribed, and B : Y → X is a bounded linear operator satisfying
BgX ≤ cB
2 gY for all g ∈ Y.
Hence we have to find u ∈ X as the solution of the variational problem
Au, v = Bg, v for all v ∈ X. (8.9)
The approximate solution uM ∈ XM is then given as in (8.3) as the unique

AuM , vM = Bg, vM for all vM ∈ XM . (8.10)
The generation of the linear system (8.5) then requires the computation of
f = Bg, ϕ = g, B ϕ for = 1, . . . , M,
i.e. we have to evaluate the application of the operator B : Y → X or of

the adjoint operator B : X → Y . In what follows we will replace the given
function g by an approximation

N
gN = gi ψi ∈ YN = span{ψi }N
i=1 ⊂ Y.
i=1
M ∈ XM of the perturbed
Then we have to find an approximate solution u
variational problem
A
uM , vM = BgN , vM for all vM ∈ XM . (8.11)
This is equivalent to the linear system
= BN g
AM u (8.12)
with matrices defined by
AM [, k] = Aϕk , ϕ , BN [, i] = Bψi , ϕ
for i = 1, . . . , N and k, = 1, . . . , M , as well as with the vector g describing

the approximation gN . The matrix BN can hereby be computed independently
of the given approximate function gN . From the X–ellipticity of the opera-
tor A we find the positive definiteness of the matrix AM , and therefore the
unique solvability of the linear system (8.12) and therefore of the equivalent
variational problem (8.11). Obviously, we have to recognize the error which is
introduced by the approximation of the given data in the linear form of the
right hand side.
8.3 Approximation of the Operator 191
Theorem 8.2 (Strang Lemma). Let A : X → X be a bounded linear

and X–elliptic operator. Let u ∈ X be the unique solution of the continuous
variational problem (8.9), and let uM ∈ XM be the unique solution of the
Galerkin variational problem (8.10). For the unique solution uM ∈ XM of the
perturbed variational problem (8.11) there holds the error estimate

1
M X ≤ A cA
u − u 2 inf u − v
M X + cB
2 g − g
N Y .
c1 vM ∈XM
Proof. When subtracting the perturbed variational problem (8.11) from the
Galerkin variational problem (8.10) this gives
A(uM − u
M ), vM = B(g − gN ), vM for all vM ∈ XM .
In particular for the test function vM := uM − u

M ∈ XM we obtain from the
X–ellipticity of A and using the boundedness of B
cA M 2X ≤ A(uM − u
1 uM − u M ), uM − u
M
= B(g − gN ), uM − u
M
≤ B(g − gN )X uM − u
M X
≤ cB
2 g − gN Y uM − u
M X .
Hence we get the estimate
cB
2
uM − u
M X ≤ g − gN Y .
cA
1
Applying the triangle inequality
u − u
M X ≤ u − uM X + uM − u
M X
we finally obtain the assertion from Theorem 8.1 (Cea’s Lemma).

8.3 Approximation of the Operator

Besides an approximation of the given right hand side we also have to con-
sider an approximation of the given operator, e.g. when applying numerical
integration schemes. Instead of the Galerkin variational problem (8.3) we then
have to find the solution uM ∈ XM of the perturbed variational problem
uM , vM = f, vM for all vM ∈ XM .
A (8.13)
: X → X is a bounded linear operator satisfying
In (8.13) A
X ≤
Av 2 vX
cA for all v ∈ X. (8.14)
Subtracting the perturbed variational problem (8.13) from the Galerkin vari-
ational problem (8.3) we find
uM , vM = 0 for all vM ∈ XM .

AuM − A (8.15)
To ensure the unique solvability of the perturbed variational problem (8.13) we

From this
have to assume the discrete stability of the approximate operator A.
we then also obtain an error estimate for the approximate solution u M ∈ XM
of (8.13).
Theorem 8.3 (Strang Lemma). Assume that the approximate operator

: X → X is XM –elliptic, i.e.
A
M , vM ≥
Av 2
1 vM X
cA for all vM ∈ XM . (8.16)
Then there exists a unique solution u M ∈ XM of the perturbed variational

problem (8.13) satisfying the error estimate
A
1 A c 1 X .
u − u
M M ≤ 1 + A (c2 + c2 ) 2A inf u − vM X + A (A − A)u
A

c1 c1 vM ∈XM
c1
(8.17)
Proof. The unique solvability of the variational problem (8.13) is a direct

consequence of the XM –ellipticity of the approximate operator A, since the

associated stiffness matrix AM is positive definite.
Let uM ∈ XM be the unique solution of the variational problem (8.3).
Using again the assumption that A is XM –elliptic, and using the orthogonality
relation (8.15) we obtain
M −u
M 2X ≤ A(u
1 uM − u
cA M ), uM − u
M
− A)uM , uM − u
= (A M
− A)uM X uM − u
≤ (A M X
and therefore
1 M X ∗ .
uM − u
M X ≤ (A − A)u

cA
1
: X → X are bounded, this gives

Since both operators A, A
M X ≤ (A − A)u
(A − A)u X + (A − A)(u
− uM )X
X + [cA
≤ (A − A)u 2 +2 ] u − uM X .
cA
Applying the triangle inequality we obtain

8.4 Galerkin–Petrov Methods 193
u − u
M X ≤ u − uM X + uM − u
M X
1 M X
≤ u − uM X + (A − A)u

cA
1
1 X + 1 [cA
≤ u − uM X + (A − A)u 2 +2 ] u − uM X
cA
A
c1
cA
1
and the assertion finally follows from Theorem 8.1 (Cea’s Lemma).

8.4 Galerkin–Petrov Methods

Let B : X → Π be a bounded linear operator, and let us assume that the
stability condition
Bv, q
cS vX ≤ sup for all v ∈ (ker B)⊥ ⊂ X (8.18)
0=q∈Π qΠ
is satisfied. Then, for a given g ∈ ImX (B) there exists a unique solution
u ∈ (ker B)⊥ of the operator equation Bu = g (cf. Theorem 3.7) satisfying
For M ∈ N we introduce two sequences of conforming trial spaces
⊥
XM = span{ϕk }M
k=1 ⊂ (ker B) , ΠM = span{ψk }M
k=1 ⊂ Π.
Using (8.2) we can define an approximate solution uM ∈ XM as the solution

of the Galerkin–Petrov variational problem
BuM , qM = g, qM for all qM ∈ ΠM . (8.19)
In contrast to the Galerkin–Bubnov variational problem (8.3) we now have
two different test and trial spaces.
Due to ΠM ⊂ Π we have the Galerkin orthogonality
B(u − uM ), qM = 0 for all qM ∈ ΠM .
The variational problem (8.19) is equivalent to the linear system BM uM = g
with the matrix BM defined by
BM [, k] = Bϕk , ψ
for k, = 1, . . . , M and with the right hand side g given by
g = g, ψ
for = 1, . . . , M . As in the continuous case we obtain the unique solvability
of the linear system when assuming a discrete stability condition,
BvM , qM
cS vM X ≤
sup for all vM ∈ XM . (8.20)
0=qM ∈ΠM qM Π
Theorem 8.4. Let u ∈ (ker B)⊥ be the unique solution of the operator equa-
tion Bu = g, and let uM ∈ XM be the unique solution of the variational prob-
lem (8.19). We further assume the discrete stability condition (8.20). Then
there holds the error estimate

cB
2
u − uM X ≤ 1 + inf u − vM X .

cB vM ∈XM
Proof. For an arbitrary v ∈ (ker B)⊥ ⊂ X there exists a uniquely determined

vM = PM v ∈ XM as the unique solution of the variational problem
BvM , qM = Bv, qM for all qM ∈ ΠM .
For the solution vM ∈ XM we obtain from the discrete stability condition

BvM , qM Bv, qM

cS vM X ≤ sup = sup ≤ cB
2 vX .
0=qM ∈ΠM qM Π 0=qM ∈ΠM qM Π
For any v ∈ (ker B)⊥ we therefore obtain a unique vM = PM v ∈ ΠM satisfying
cB
2
PM vX ≤ vX .

cS
In particular, for the unique solution uM ∈ XM of the variational problem
(8.19) we obtain uM = PM u. On the other hand we have vM = PM vM for all
vM ∈ XM . Hence we have for an arbitrary vM ∈ XM
u − uM X = u − vM + vM − uM X = u − vM − PM (u − vM )X

cB
≤ u − vM X + PM (u − vM )X ≤ 1 + 2 u − vM X

cS
and therefore the assertion.
The convergence of the approximate solution uM → u ∈ X as M → ∞
then follows as for a Galerkin–Bubnov method from an approximation prop-
erty of the trial space XM .
It remains to establish the discrete stability condition (8.20). A possible
criterion is the following result due to Fortin [58].
Lemma 8.5 (Criterium of Fortin). Let B : X → Π be a bounded linear

operator, and let the continuous stability condition (8.18) be satisfied. If there
exist a bounded projection operator RM : Π → ΠM satisfying
BvM , q − RM q = 0 for all vM ∈ XM
and
RM qΠ ≤ cR qΠ for all q ∈ Π,
then there holds the discrete stability condition (8.20) with
cS = cS /cR .
Proof. Using the stability condition (8.18) we have for qN ∈ ΠN ⊂ Π
BvM , q BvM , RM q
cS vM X ≤ sup = sup
0=q∈Π qΠ 0=q∈Π qΠ
BvM , RM q BvM , qM
≤ cR sup ≤ cR sup ,
0=q∈Π RM qΠ 0=qM ∈ΠM qM Π
and therefore the discrete stability condition (8.20).

8.5 Mixed Formulations

Now we consider the approximate solution of the saddle point problem (3.22)
to find (u, p) ∈ X × Π such that
Au, v + Bv, p = f, v

(8.21)
Bu, q = g, q
is satisfied for all (v, q) ∈ X × Π.

We assume that A : X → X and B : X → Π are bounded linear oper-
ators, and that A is X–elliptic. For example, the last assumption is satisfied
when considering the Stokes system and the modified variational formula-
tion (4.22) and (4.23) for a Dirichlet boundary value problem with Lagrange
multipliers. We further assume the stability condition (8.18). Hence, all as-
sumptions of Theorem 3.11 and of Theorem 3.13 are satisfied, and there exists
a unique solution (u, p) ∈ X × Π of the saddle point problem (8.21).
For N, M ∈ N we define two sequences of conforming trial spaces
XM = span{ϕk }M
k=1 ⊂ X, ΠN = span{ψi }N
i=1 ⊂ Π.
Then the Galerkin variational formulation of the saddle point problem (8.21)
is to find (uM , pN ) ∈ XM × ΠN such that
AuM , vM + BvM , pN = f, vM

(8.22)
BuM , qN = g, qN
is satisfied for all (vM , qN ) ∈ XM × ΠN . With the matrices AM and BN

defined by
AM [, k] = Aϕk , ϕ , BN [j, k] = Bϕk , ψj
for k, = 1, . . . , M , j = 1, . . . , N , and with the vectors f and g given by
f = f, ϕ , gj = g, ψj
for = 1, . . . , M , j = 1, . . . , N , the variational formulation (8.22) is equivalent

to the linear system
& '& ' & '
u f
AM BN
= . (8.23)
BN 0 p g
We first consider the unique solvability of the linear system (8.23) from an
algebraic point of view. The dimension of the system matrix K in (8.23) is
N + M . With
rang AM ≤ M, rang BN ≤ min{M, N }
we find
rang K ≤ M + min{M, N } .
In particular for M < N we obtain
rang K ≤ 2M < M + N = dim K.
i.e. the linear system (8.23) is in general not solvable. Hence we have to define
the trial spaces XM and ΠN with care. The necessary condition M ≥ N
shows, that the trial space XM has to be rich enough compared with the trial
space ΠN .
To investigate the unique solvability of the Galerkin variational problem
(8.22) we will make use of Theorem 3.13. When considering the conform-
ing trial space XM ⊂ X we obtain from the X–ellipticity of A the positive
definiteness of the matrix AM , i.e.
2
(AM v, v) = AvM , vM ≥ cA
1 vM X > 0
for all 0 = v ∈ RM ↔ vM ∈ XM . The matrix AM is therefore invertible

and the linear system (8.23) can be transformed into the Schur complement
system
BN A−1 −1
M BN p = BN AM f − g . (8.24)
It remains to investigate the unique solvability of the linear system (8.24).
For this we assume the discrete stability or Babus̆ka–Brezzi–Ladyshenskaya
(BBL) condition
BvM , qN
cS qN Π ≤
sup for all qN ∈ ΠN . (8.25)
0=vM ∈XM vM X
It is worth to remark, that the discrete stability condition (8.25) is in general
not an immediate consequence of the continuous stability condition (3.25).
Lemma 8.6. Let A : X → X and B : X → Π be bounded linear operators,
and let A be X–elliptic. For the conforming trial spaces XM ⊂ X and ΠN ⊂ Π
we assume the discrete stability condition (8.25). The symmetric Schur com-
plement matrix SN := BN A−1
M BN of the Schur complement system (8.24) is
then positive definite, i.e.
(SN q, q) ≥ cS1 N qN 2Π
for all 0 = q ∈ ΠN ↔ qN ∈ ΠN .
Proof. For an arbitrary but fixed q ∈ RN we define ū := A−1

M BN q, i.e. for the
associated functions qN ∈ ΠN and ūM ∈ XM we have
AūM , vM = BvM , qN for all vM ∈ XM .
Using the X–ellipticity of A we then obtain

2 −1
1 ūM X ≤ AūM , ūM = B ūM , qN = (BN ū, q) = (BN AM BN q, q).
cA
On the other hand, the discrete stability condition (8.25) gives
BvM , qN AūM , vM
cS qN Π ≤ sup = sup ≤ cA
2 ūM X
0=v M ∈RM vM X 0=v M ∈RM vM X
and with
2 2
cA
2 1 cA
2
qN 2Π ≤ ūM 2X ≤ A (BN A−1
M BN q, q)
cS c1 cS
we finally get the assertion.

Hence we have the unique solvability of the Schur complement system
(8.24) and therefore of the linear system (8.23). Moreover, we also have the
following stability estimate.
Theorem 8.7. Let A : X → X and B : X → Π be bounded and linear

operators, and let A be X–elliptic. For the conforming trial spaces XM ⊂ X
and ΠN ⊂ Π we assume the discrete stability condition (8.25). For the unique
solution (uM , pN ) ∈ XM ×ΠN of the saddle point problem (8.22) we then have
the stability estimates
1 cB2 1
pN Π ≤ SN cA
f X + SN gΠ (8.26)
c1 1 c1
and
cB cB 1 cB
uM X ≤ 1 + S2N 2A f X + SN 2A gΠ . (8.27)
c1 c1 c1 c1
Proof. Let (u, p) ∈ RM × RN ↔ (uM , pN ) ∈ XM × ΠN be the unique solution

of the linear system (8.23) and of the saddle point problem (8.22), respectively.
Using Lemma 8.6 we have
cS1 N pN 2Π ≤ (SN p, p) = (BN A−1 −1

M BN p, p) = (BN AM f − g, p)
0 B 1
= B ūM − g, pN ≤ c2 ūM X + gΠ pN Π
and therefore
1 0 B 1
pN Π ≤ c2 ūM X + gΠ .
cS1 N
Hereby, ū = A−1
M f ∈R
M
↔ ūM ∈ XM is the unique solution of the variational
problem
AūM , vM = f, vM for all vM ∈ XM .
From the X–ellipticity of A we then find
1
ūM X ≤ f X .
cA
1
Moreover,
2
1 uM X ≤ AuM , uM
cA
0 1
2 pN Π uM X
= f, uM − BuM , pN ≤ f X + cB
and therefore
1 0 1
uM X ≤ A
f X + cB
2 pN Π .
c1
With (8.26) we then obtain (8.27) .
Using the stability estimates (8.26) and (8.27) we also obtain an error
estimate for the approximate solution (uM , pN ) ∈ XM × ΠN .
Theorem 8.8. Let all assumptions of Theorem 8.7 be valid. For the unique
approximate solution (uM , pN ) ∈ XM × ΠN of the saddle point problem (8.22)

u − uM X + p − pN Π ≤ c inf u − vM X + inf p − qN Π .
vM ∈XM qN ∈ΠN
Proof. When taking the difference of the continuous saddle point formulation
(8.21) with the Galerkin variational problem (8.22) for the conforming trial
spaces XM × ΠN ⊂ X × Π we obtain the Galerkin orthogonalities
A(u − uM ), vM + BvM , p − pN = 0
B(u − uM ), qN =0
for all (vM , qN ) ∈ XM × ΠN . For arbitrary (ūM , p̄N ) ∈ XM × ΠN we then

obtain
A(ūM − uM ), vM + BvM , p̄N − pN = A(ūM − u) + B (p̄N − p), vM
B(ūM − uM ), qN = B(ūM − u), qN
for all (vM , qN ) ∈ XM × ΠN . Using Theorem 8.7 we find the unique solution
(ūM − uM , p̄N − pM ) ∈ XM × ΠN , and we obtain the stability estimates
p̄N − pN Π ≤ c1 A(ūM − u) + B (p̄N − p)X + c2 B(ūM − u)Π ,

ūM − uM X ≤ c3 A(ūM − u) + B (p̄N − p)X + c4 B(ūM − u)Π
for arbitrary (ūM , p̄N ) ∈ XM × ΠN . Due to the mapping properties of the

bounded operators A, B and B we get with the triangle inequality
p − pN Π ≤ p − p̄N Π + p̄N + pN Π
≤ (1 + c1 cB
2 ) p − p̄N Π + (c1 c2 + c2 c2 ) u − ūM X
A B
for arbitrary (ūM , p̄N ) ∈ XM × ΠN . The estimate for u − uM X follows in

the same way.
It remains to validate the discrete stability condition (8.25). As in Lemma
8.5 we can use the criterion of Fortin to establish (8.25).
Lemma 8.9 (Criteria of Fortin). Let B : X → Π be a bounded linear

operator, and let the continuous stability condition (3.25) be satisfied. If there
exists a bounded projection operator PM : X → XM satisfying
B(v − PM v), qN = 0 for all qN ∈ ΠN
and
PM vX ≤ cP vX for all v ∈ X,
then there holds the discrete stability condition (8.25) with
cS = cS /cP .
8.6 Coercive Operators
We finally consider an approximate solution of the operator equation Au = f

when A : X → X is assumed to be coercive, i.e. there exists a compact
operator C : X → X such that Gårding’s inequality (3.32) is satisfied,
2
(A + C)v, v ≥ cA
For a sequence XM ⊂ X of finite dimensional trial spaces we consider the

Galerkin variational problem to find uM ∈ XM such that
AuM , vM = f, vM (8.28)
is satisfied for all vM ∈ XM . Note that the variational problem (8.28) for-
mally coincides with the Galerkin–Bubnov formulation (8.3). However, since
we now consider the more general case of a coercive operator instead of an el-
liptic operator, the numerical analysis to establish suitable stability and error
estimates is different.
Theorem 8.10 (Cea’s Lemma). Let A : X → XM be a bounded linear

coercive operator and let the stability condition
AwM , vM
cS wM X ≤ sup (8.29)
vM ∈XM ,
vM
X >0 vM X
be satisfied for all wM ∈ XM . Then there exists a unique solution uM ∈ XM

of the Galerkin variational problem (8.28) satisfying the stability estimate
1
uM X ≤ f X (8.30)
cS
and the error estimate
9 cA :
2
u − uM X ≤ 1+ inf u − vM X . (8.31)
cS vM ∈XM
Proof. We consider the homogeneous linear system AM w̄ = 0 to find an ap-

proximate solution w̄M ∈ XM of the homogeneous variational problem
Aw̄M , vM = 0 for all vM ∈ XM .
Using the stability condition (8.29) we then obtain
Aw̄M , vM
cS w̄M X ≤ sup = 0
vM ∈XM ,
vM
X >0 vM X
and therefore w̄M = 0 ↔ w̄ = 0. This ensures the unique solvability of the

linear system AM u = f and therefore of the variational problem (8.28).
Let u ∈ RM ↔ uM ∈ XM be the unique solution of the Galerkin varia-
tional problem (8.28). Again, applying the stability condition (8.29) this gives
AuM , vM
cS uM X ≤ sup
vM ∈XM ,
vM
X >0 vM X
f, vM
= sup ≤ f X
vM ∈XM ,
vM
X >0 vM X
and therefore the stability estimate (8.30).

For an arbitrary w ∈ X we define an approximate solution wM ∈ XM of
the Galerkin variational formulation
AwM , vM = Aw, vM for all vM ∈ XM .
This defines the projection operator GM : X → XM by wM = GM w satisfying
1 cA
GM wX = wM X ≤ AwX ≤ 2 wX .
cS cS
In particular, we have uM = GM u for the solution of the Galerkin variational
formulation (8.28). Since GM is a projection, GM vM = vM for all vM ∈ XM ,
we then find
u − uM X = u − vM + GM vM − uM X
9 cA :
2
≤ u − vM X + GM (u − vM )X ≤ 1+ u − vM X
cS
for all vM ∈ XM . From this, the error estimate (8.31) follows.

It remains to validate the discrete stability condition (8.29). Note, that for
an X–elliptic operator A we then obtain
AwM , wM AwM , vM
1 wM X ≤
cA ≤ sup
wM X vM ∈XM ,
vM
X >0 vM X
for all wM ∈ XM , i.e. (8.29). In what follows we consider the case of an

coercive operator A.
Theorem 8.11. Let A : X → X be a bounded linear operator which is as-

sumed to be coercive and injective. Let XM ⊂ X be a dense sequence of con-
forming trial spaces. Then there exists an index M0 ∈ N such that the discrete
stability condition (8.29) is satisfied for M ≥ M0 .
Proof. Let wM ∈ XM be arbitrary but fixed. Since A : X → X is assumed to

be coercive, there is a compact operator C : X → X such that the bounded
operator D = A + C : X → X is X–elliptic. Hence we can set v̄ = D−1 CwM
as the unique solution v̄ ∈ X of the variational problem
Dv̄, v = CwM , v for all v ∈ X.
Moreover we can define an approximate solution v̄M ∈ XM as the unique

solution of the Galerkin variational problem
Dv̄M , vM = CwM , vM for all vM ∈ XM .
Hence we have the Galerkin orthogonality
D(v̄ − v̄M ), vM = 0 for all vM ∈ XM .
Applying Cea’s lemma (cf. Theorem 8.1) for the X–elliptic operator D we
also find the stability estimate
1 cC
2
v̄M X ≤ D
CwM X ≤ D wM X
c1 c1
and therefore
9 cC :
2
wM − v̄M X ≤ wM X + vM X ≤ 1+ D
wM X
c1

cD
2
v̄ − v̄M X ≤ inf v̄ − vM X .
cD
1 vM ∈XM
Hence, the approximation property (8.8) of the trial space gives the conver-
gence v̄M ∈ v̄ in X for M → ∞.
Considering as test function vM = wM − v̄M we obtain
AwM , wM − v̄M = AwM , wM − v̄ + AwM , v̄ − v̄M

= AwM , wM − D−1 CwM + AwM , v̄ − v̄M .
For the first summand we further get
AwM , wM − D−1 CwM = AwM , D−1 (D − C)wM

= AwM , D−1 AwM
2 2
≥ cD
1 AwM X ≥ c1 cA wM X
D
since A : X → X has a bounded inverse. On the other hand, using the

Galerkin orthogonality we get
|AwM , v̄ − v̄M | = |DwM , v̄ − v̄M − CwM , v̄ − v̄M |

= |wM , D(v̄ − v̄M ) − wM , C(v̄ − v̄M )|
= |wM , C(v̄ − v̄M )| ≤ wM X C(v̄ − v̄M )X
Since C : X → X is a compact operator, there exists a subsequence {v̄M }M ∈N

satisfying
C(v̄ − v̄M )X
lim = 0.
M →∞ v̄X
Hence there exists an index M0 ∈ N such that
1 D
AwM , wM − v̄M ≥ c cA wM 2X
2 1
1 9 cD :−1
2
≥ cD1 cA 1 + D wM X wM − v̄M X
2 c1
is satisfied for all M ≥ M0 which implies the stability condition (8.29).

8.7 Exercises
8.1 Let X be a Hilbert space and let a(·, ·) : X × X → R be a symmetric
and positive definite bilinear form. For the approximation of the minimization
problem
1
F (u) = min F (v), F (v) = a(v, v) − f, v
v∈X 2
we introduce a finite–dimensional trial space
XM = span{ϕk }M
k=1 ⊂ X.
Derive the variational problem to find the approximate solution uM ∈ XM .

9
Finite Elements
For the approximate solution of variational formulations as described in

Chapter 4 we introduce appropriate finite–dimensional trial spaces, and prove
certain approximation properties in Sobolev spaces. For simplicity we just
consider lowest order polynomial basis functions. For an introduction of more
general finite elements we refer, for example, to [31, 41, 85].
9.1 Reference Elements

Let Ω ⊂ Rd (d = 1, 2, 3) be a bounded domain with a polygonal (d = 2)
or with a polyhedral (d = 3) boundary. We consider a sequence {TN }N ∈N of
decompositions (meshes)
*
N
Ω = TN = τ (9.1)
=1
with finite elements τ . In the simplest case we have an interval (d = 1), a

triangle (d = 2), or a tetrahedron (d = 3). Further we denote by {xk }M k=1 the
set of all nodes of the decomposition TN , see Fig. 9.1 for a finite element τ
and the corresponding nodes xk . In addition, for d = 2, 3 we have by {kj }K j=1
the set of all edges.
By I(k) we denote the index set of all elements τ where xk ∈ τ is a node,
I(k) := { ∈ N : xk ∈ τ } for k = 1, . . . , M.
Moreover,
J() := {k ∈ N : xk ∈ τ } for = 1, . . . , N
is the index set of all nodes xk with xk ∈ τ . Note that dim J() = d + 1 in
the case of the finite elements τ considered here. Finally,
K(j) := { ∈ N : kj ∈ τ } for j = 1, . . . , K
204 9 Finite Elements
r r
@ @
JJ@
@
@ J@
r r J @
@
@ Jr @
HH
r @r r
H@r
d=1 d=2 d=3
Fig. 9.1. Finite element τ and related nodes xk .
is the index set of all elements τ with the edge kj .

The decomposition (9.1) is called admissible, if two neighboring elements
join either a node (d = 1, 2, 3), an edge (d = 2, 3), or a triangle (d = 3),
see Fig. 9.2. In particular, we avoid hanging nodes as in the inadmissible
decomposition of Fig. 9.2.
t t
HH HH
HH HH
Ht Ht

t

S

tP P

t S
PP PP S
PP PP S
PPt PPSt
Fig. 9.2. Admissible and inadmissible triangulations (d = 2).
In what follows we only consider admissible decompositions of the computa-

tional domain Ω. For a finite element τ

∆ := dx (9.2)
τ
is the volume, while

1/d
h := ∆
is the local mesh size. Moreover,
d := sup |x − y|
x,y∈τ
is the diameter of the finite element τ , which coincides with the longest edge
of the element τ . Obviously, for d = 1 we have
∆ = h = d .
9.1 Reference Elements 205
Finally, r is the radius of the largest circle (d = 2) or sphere (d = 3) which can

be inscribed in the finite element τ . A finite element τ of the decomposition
(9.1) is called shape regular, if the diameter d of the finite element τ is
bounded uniformly by the radius r , i.e.
d ≤ cF r for = 1, . . . , N
where the constant cF does not depend on TN . For the two–dimensional case
d = 2 we then have
πr 2 ≤ ∆ = h2 ≤ d2 ≤ c2F r 2 ,
and therefore the equivalence relations

√
π r ≤ h ≤ d ≤ cF r .
Correspondingly, for d = 3 we have

4 3
πr ≤ ∆ = h3 ≤ d3 ≤ c3F r 3
3
and therefore 8
34
π r ≤ h ≤ d ≤ cF r .
3
The global mesh size h is defined by
h = hmax := max h
=1,...,N
while
hmin := min h
=1,...,N
is the minimal local mesh size. The family of decompositions TN is called

globally quasi–uniform, if
hmax
≤ cG
hmin
is bounded by a global constant cG ≥ 1 which is independent of N ∈ N. The
family TN is called locally quasi–uniform, if
h
≤ cL for = 1, . . . , N
hj
holds for all neighboring finite elements τj and τ . Here, two finite elements τ
and τj are called neighboring, if the average τ ∩ τ j consists either of a node,
an edge, or a triangle.
In the one–dimensional case d = 1 each finite element τ can be described
via a local parametrization, in particular for x ∈ τ and 1 , 2 ∈ J() we have
x = x 1 + ξ (x 2 − x 1 ) = x 1 + ξ h for ξ ∈ (0, 1).

Here, the element

τ := (0, 1) (9.3)
is called reference element. If we consider a function v(x) for x ∈ τ we can
write
v(x) = v(x 1 + ξ h ) =: v (ξ) for ξ ∈ τ,
in particular we can identify a function v(x) for x ∈ τ with a function in the
reference element, v (ξ) for ξ ∈ τ . It follows that

v2L2 (τ ) = |v(x)|2 dx = v (ξ)|2 h dξ = h
| v 2L2 (τ ) .
τ τ
For the first derivative we have, by applying the chain rule,

d d
v (ξ) = h v(x) for x ∈ τ , ξ ∈ τ
dξ dx
and therefore
d 1 d
v(x) = v (ξ) for x ∈ τ , ξ ∈ τ.
dx h dξ
For m ∈ N the recursive application of this result gives
dm −m d
m
v(x) = h v (ξ) for x ∈ τ , ξ ∈ τ.
dxm dξ m
Hence we obtain for the local norms of v and v

m 2 m 2
d
1−2m d

dxm v = h dξ m v for m ∈ N0 . (9.4)
L2 (τ ) L2 (τ )
In the two–dimensional case d = 2 the reference element τ is given by the

triangle
τ = ξ ∈ R2 : 0 ≤ ξ1 ≤ 1, 0 ≤ ξ2 ≤ 1 − ξ1 . (9.5)
Then we find the local parametrization for x ∈ τ as
2

x = x 1 + ξi (x i+1 − x 1 ) = x 1 + J ξ for ξ ∈ τ
i=1
with the Jacobian

& '
x 2 ,1 − x 1 ,1 x 3 ,1 − x 1 ,1
J = .
x 2 ,2 − x 1 ,2 x 3 ,2 − x 1 ,2
To compute the area (volume) of the finite element τ we obtain

ξ2
6
x3
1
C @
C x(ξ) = x1 + J ξ @
k6 C @
C @
C @
τ k
C 5 @
C τ @
xP PPP C
1
PP @
k4 PP C @@ -
P
PC x 0 1 ξ1
2
Fig. 9.3. Finite element τ and reference element τ (d = 2).
1 1−ξ
1
1
∆ = dsx = |det J | dξ = |det J | dξ2 dξ1 = |det J |
2
τ τ 0 0
and therefore
|detJ | = 2 ∆ . (9.6)
If we consider a function v(x) for x ∈ τ we can write
v(x) = v(x 1 + J ξ) = v (ξ) for ξ ∈ τ.
Then, by applying the chain rule we get
∇ξ v (ξ) = J ∇x v(x)
and therefore
∇x v(x) = J − ∇ξ v (ξ).
As for the one–dimensional case d = 1 we can show the following norm equiv-
alence estimates:
Lemma 9.1. For d = 2 and m ∈ N0 there hold the norm equivalence inequal-
ities
1
(2∆ )1−m ∇m 2L2 (τ ) ≤ ∇m
ξ v
2
x vL2 (τ ) ≤ cm (2∆ )
1−m
∇m 2L2 (τ )
ξ v
cm
(9.7)
where 2 m
cF
cm = .
π
Proof. For m = 0 the assertion follows directly from

v2L2 (τ ) = |v(x)|2 dx = v (ξ)|2 |detJ | dξ = 2 ∆
| v 2L2 (τ ) .
τ τ
Now we consider the case m = 1. Then,

∇x v2L2 (τ ) = |∇x v(x)|2 dx = |J − ∇ξ v(ξ)|2 |detJ | dξ
τ τ

= 2∆ (J −1 J − ∇ξ v(ξ), ∇ξ v(ξ)) dξ
τ

≤ 2∆ λmax (J −1 J − ) |∇ξ v(ξ)|2 dξ
τ
= 2∆ λmax (J −1 J − ) ∇ξ v2L2 (τ )
as well as
∇x v2L2 (τ ) ≥ 2∆ λmin (J −1 J − ) ∇ξ v2L2 (τ ) .
It is therefore sufficient to estimate the eigenvalues of the matrix J J . With
a := |x 2 − x 1 |, b := |x 3 − x 1 |, α = <)(x 3 − x 1 , x 2 − x 1 )
we have & '

a2 a b cos α
J J = ,
a b cos α b2
and the eigenvalues of J J are

1 2 2 . 3
λ1/2 = a + b2 ± (a2 − b2 )2 + 4a2 b2 cos2 α .
2
Obviously, for the maximal eigenvalue λ1 we have the inclusion
1 2
(a + b2 ) ≤ λ1 ≤ a2 + b2
2
while for the product of the eigenvalues λ1/2 we have with (9.6)
λ1 λ2 = det(J J ) = |det J |2 = 4∆2 .
The minimal eigenvalue λ2 admits the lower estimate
4∆2 4∆2
λ2 = ≥ 2 2
λ1 a +b
and therefore we conclude
4∆2
≤ λmin (J J ) ≤ λmax (J J ) ≤ a2 + b2 .
a2 + b2
Moreover,
2c2F
a2 + b2 ≤ 2 d2 ≤ 2 c2F r 2 ≤ ∆ .
π
Hence we have
2π 2c2
2 ∆ ≤ λmin (J J ) ≤ λmax (J J ) ≤ F ∆
cF π
and the eigenvalues of the inverse matrix J −1 J − can be estimated as
π c2
2 (2∆ )−1 ≤ λmin (J −1 J − ) ≤ λmax (J −1 J − ) ≤ F (2∆ )−1 .
cF π
Hence we conclude the norm equivalence inequalities for m = 1. For m > 1,
the assertion follows by recursive applications of the above estimates.

In the three–dimensional case d = 3 the reference element τ is given by
the tetrahedron

τ = ξ ∈ R3 : 0 ≤ ξ1 ≤ 1, 0 ≤ ξ2 ≤ 1 − ξ1 , 0 ≤ ξ3 ≤ 1 − ξ1 − ξ2 . (9.8)
For x ∈ τ we then have the local parametrization
3

x = x 1 + ξi (x i+1 − x 1 ) = x 1 + J ξ for ξ ∈ τ
i=1
with the Jacobian

⎛ ⎞
x 2 ,1 − x 1 ,1 x 3 ,1 − x 1 ,1 x 4 ,1 − x 1 ,1
⎜ ⎟
J = ⎝ x 2 ,2 − x 1 ,2 x 3 ,2 − x 1 ,2 x 4 ,2 − x 1 ,2 ⎠ .
x 2 ,3 − x 1 ,3 x 3 ,3 − x 1 ,3 x 4 ,3 − x 1 ,3
For the volume of the finite element τ we find

∆ = dsx = |det J | dξ
τ τ
1 1−ξ
1 1−ξ
1 −ξ2
1
= |det J | dξ3 dξ2 dξ1 = |det J | (9.9)
6
0 0 0
and therefore
|det J | = 6 ∆ . (9.10)
As for the two–dimensional case we can write a function v(x) for x ∈ τ as
v(x) = v(x 1 + J ξ) = v (ξ) for ξ ∈ τ.
Again, the application of the chain rule gives
∇ξ v (ξ) = J ∇x v(x), ∇x v(x) = J − ∇ξ v (ξ)
and in analogy to Lemma 9.1 we have:
Lemma 9.2. For d = 3 and m ∈ N0 there hold the norm equivalence inequal-
ities
c1 ∆ h−2m
∇m 2L2 (τ ) ≤ ∇m
ξ v
2 −2m
x vL2 (τ ) ≤ c2 ∆ h ∇m 2L2 (τ )
ξ v
with positive constants c1 and c2 which may depend on m and on cF .

Proof. For m = 0 a direct computation gives

v(x)2L2 (τ ) = |v(x)|2 dx = v (ξ)|2 |det J | dξ = 6∆
| v 2L2 (τ ) .
τ τ
For m = 1 we first obtain, as in the proof of Lemma 9.1, the equivalence

inequalities
6∆ λmin (J −1 J − ) ∇ξ v 2L2 (τ ) ≤ ∇x v2L2 (τ )

≤ 6∆ λmax (J −1 J − ) ∇ξ v 2L2 (τ ) .
Hence we have to estimate the eigenvalues of the symmetric and positive

definite matrix ⎛ ⎞
a2 ab cos α ac cos β
⎜ ⎟
J J = ⎝ ab cos α b2 bc cos γ ⎠
ac cos β bc cos γ c2
where
a := |x 2 − x 1 |, b := |x 3 − x 1 |, c := |x 4 − x 1 |
and
α := <)(x 2 − x 1 , x 3 − x 1 ),
β := <)(x 2 − x 1 , x 4 − x 1 ),
γ := <)(x 3 − x 1 , x 4 − x 1 ).
From 0 < λi for i = 1, 2, 3 and
λ1 + λ2 + λ3 = a2 + b2 + c2
we can estimate the maximal eigenvalue by
λmax (J J ) ≤ a2 + b2 + c2 .
The product of all eigenvalues can be written by using (9.10) as
λ1 λ2 λ3 = det(J J ) = |det J |2 = 36 ∆2
and hence we obtain an estimate for the minimal eigenvalue

36∆2 36∆2
λmin (J J ) ≥
≥ 2 .
[λmax (J J )]2 [a + b2 + c2 ]2
9.2 Form Functions 211
Altogether we therefore have
36∆2
≤ λmin (J J ) ≤ λmax (J J ) ≤ a2 + b2 + c2 .
[a2 + b2 + c2 ]2
Since the finite element τ is assumed to be shape regular, we can estimate
the length of all edges by
8
2 2 2 2 2 2 3 9 2 2
a + b + c ≤ 3d ≤ 3cF r ≤ 3 c h
16π 2 F
and hence we obtain
8 8
4 3 256π 4 2 3 9 2 2
h ≤ λmin (J J ) ≤ λmax (J J ) ≤ 3 c h .
c4F 81 16π 2 F
The assertion now follows as in the proof of Lemma 9.1.

By using the norm equivalence estimates (9.4) for d = 1 as well as Lemma
9.1 for d = 2 and Lemma 9.2 for d = 3 we can formulate the following result:
Theorem 9.3. Let τ ⊂ Rd be a finite element of a shape regular and admis-

sible decomposition TN . If v is sufficiently smooth we then have for m ∈ N0
c1 ∆ h−2m
∇m 2L2 (τ ) ≤ ∇m
ξ v
2 −2m
x vL2 (τ ) ≤ c2 ∆ h ∇m 2L2 (τ )
ξ v
with positive constants c1 and c2 which may depend on m and on cF .
9.2 Form Functions

With respect to the decomposition TN as defined in (9.1) we now introduce
trial spaces of piecewise polynomial functions. The related basis functions,
which are associated to global degrees of freedom, are defined locally by using
suitable form functions which are formulated with respect to an element τ .
We consider a reference element τ which is either an interval (9.3) for
d = 1, a triangle (9.5) for d = 2, or a tetrahedron (9.8) for d = 3.
The simplest form functions are the constant functions
ψ10 (ξ) = 1 for ξ ∈ τ.
If we consider a function vh (x) which is constant for x ∈ τ we then have the

representation
vh (x) = vh (x 1 + J ξ) = v ψ10 (ξ) for x ∈ τ , ξ ∈ τ, (9.11)
where v is the associated coefficient describing the value of vh on τ . Moreover,

we have
vh 2L2 (τ ) = ∆ v 2 . (9.12)
If we consider a function vh (x) which is linear for x ∈ τ , then this function is

uniquely determined by the values vk at the nodes of the reference element τ ,

d+1
vh (ξ) = vk ψk1 (ξ) for ξ ∈ τ. (9.13)
k=1
Here, the linear form functions are given for d = 1
ψ11 (ξ) := 1 − ξ, ψ21 (ξ) := ξ,
for d = 2
ψ11 (ξ) := 1 − ξ1 − ξ2 , ψ21 (ξ) := ξ1 , ψ31 (ξ) := ξ2 ,
and for d = 3
ψ11 (ξ) := 1 − ξ1 − ξ2 − ξ3 , ψ21 (ξ) := ξ1 , ψ31 (ξ) := ξ2 , ψ41 (ξ) := ξ3 .
Let τ be an arbitrary finite element with nodes x k , k ∈ J(). If vh is a linear

function on τ , then we can write

d+1
vh (x) = vh (x 1 + J ξ) = v k ψk1 (ξ) for x ∈ τ , ξ ∈ τ. (9.14)
k=1
As in (9.12) we can estimate the L2 norm vh L2 (τ ) by the Euclidean norm

of the nodal values.
Lemma 9.4. Let vh be a linear function as given in (9.14). Then,
∆
d+1
∆ 2
d+1
v 2k ≤ vh 2L2 (τ ) ≤ v k .
(d + 1)(d + 2) d+1
k=1 k=1
Proof. We can compute the local L2 norm of the linear function vh as

d+1
d+1
vh 2L2 (τ ) = vh , vh L2 (τ ) = vi vj ψi (ξ)ψj (ξ)|detJ |dξ = (G v , v )
i=1 j=1 τ
where
∆
G = (Id+1 + ed+1 e
d+1 )
(d + 1)(d + 2)
is the local mass matrix and ed+1 = 1 ∈ Rd+1 . From the eigenvalues of the
matrix Id+1 + ed+1 e
d+1 ,
λ1 = d + 2, λ2 = · · · = λd+1 = 1,
the assertion follows.

Corollary 9.5. Let vh be a linear function as given in (9.14). Then,

∆
vh 2L∞ (τ ) ≤ vh 2L2 (τ ) ≤ ∆ vh 2L∞ (τ ) .
(d + 1)(d + 2)
Proof. Obviously, the maximal value of vh and therefore the maximum norm
vh L∞ (τ ) is equal some nodal value vk∗ = vh (xk∗ ). The assertion then follows
from Lemma 9.4.
In many applications it is essential to bound the norm of the gradient of
a piecewise polynomial function by the norm of this function itself.
Lemma 9.6. Let vh be a linear function as given in (9.14). Then there holds
the local inverse inequality
∇x vh L2 (τ ) ≤ cI h−1
vh L2 (τ ) (9.15)
where cI is some positive constant.

Proof. The application of Theorem 9.3 gives first
∇x vh 2L2 (τ ) ≤ c2 ∆ h−2 2L2 (τ ) .

∇ξ v
To compute the gradient of the linear function

d+1
v (ξ) = v k ψk1 (ξ)
k=1
we obtain for d = 1
∇ξ v = v 2 − v 1
and therefore
∇ξ v 2L2 (τ ) = (v 2 − v 1 )2 ≤ 2 [v 21 + v 22 ] ≤ 4 vh 2L∞ (τ ) .
In the two–dimensional case d = 2 the gradient is

v 2 − v 1
∇ξ v =
v 3 − v 1
and therefore we obtain

10 1
∇ξ v 2L2 (τ ) = (v 2 − v 1 )2 + (v 3 − v 1 )2
2
10 2 1
≤ 2v 2 + 2v 23 + 4v 21 ≤ 4 vh 2L∞ (τ ) .
2
Finally, for d = 3 we have
⎛ ⎞
v 2 − v 1
∇ξ v = ⎝ v 3 − v 1 ⎠
v 4 − v 1
and thus
10 1
∇ξ v 2L2 (τ ) = (v 2 − v 1 )2 + (v 3 − v 1 )2 + (v 4 − v 1 )2
6
10 2 1
≤ 2v 2 + 2v 23 + 2v 24 + 6v 21 ≤ 2 vh 2L∞ (τ ) .
6
∇x vh 2L2 (τ ) ≤ 4 c2 ∆ h−2 2

vh L∞ (τ )
and the inverse inequality now follows from Corollary 9.5.

Form functions of locally higher polynomial degree can be defined hierar-
chically based on piecewise linear form functions. We define quadratic form
functions for d = 1 by
ψ12 (ξ) = 1 − ξ, ψ22 (ξ) = ξ, ψ32 (ξ) = 4ξ(1 − ξ) ,
for d = 2 by
ψ12 (ξ) = 1 − ξ1 − ξ2 , ψ22 (ξ) = ξ1 , ψ32 (ξ) = ξ2 ,

ψ42 (ξ) = 4ξ1 (1 − ξ1 − ξ2 ), ψ52 (ξ) = 4ξ1 ξ2 , ψ62 (ξ) = 4ξ2 (1 − ξ1 − ξ2 ),
and for d = 3 by
ψ12 (ξ) = 1 − ξ1 − ξ2 − ξ3 , ψ52 (ξ) = 4ξ1 (1 − ξ1 − ξ2 − ξ3 ),

ψ22 (ξ) = ξ1 , ψ62 (ξ) = 4ξ1 ξ2 ,
ψ32 (ξ) = ξ2 , ψ72 (ξ) = 4ξ2 (1 − ξ1 − ξ2 − ξ3 ),
ψ42 (ξ) = ξ3 , ψ82 (ξ) = 4ξ3 (1 − ξ1 − ξ2 − ξ3 ),
ψ92 (ξ) = 4ξ3 ξ1 ,
2
ψ10 (ξ) = 4ξ2 ξ3 .
Note that linear form functions are associated to degrees of freedom at the
nodes xk ∈ τ , while the quadratic form functions are associated to the edge
mid points x∗kj . If the function vh is quadratic on τ then we can write
1
2 (d+1)(d+2)

vh (x) = vh (x 1 + J ξ) = v k ψk2 (ξ) for x ∈ τ , ξ ∈ τ. (9.16)
k=1
As in the proof of Lemma 9.4 we have

1
2 (d+1)(d+2)

vh 2L2 (τ ) = vi vj ψi2 (ξ)ψj2 (ξ)|detJ |dξ = (G v , v ) .
i,j=1 τ
In particular for d = 1 the local mass matrix is

⎛ ⎞
1/3 1/6 1/3
G = ∆ ⎝ 1/6 1/3 1/3 ⎠
1/3 1/3 8/15
where the eigenvalues of G are

# √ $
∆ 31 89
λ1 = , λ2/3 = ∆ ± .
6 60 20
By using a similar approach in the two–dimensional case d = 2 as well as in

the three–dimensional case d = 3 we can prove equivalence estimates which
correspond to the results of Lemma 9.4,
1 1
2 (d+1)(d+2)
2 (d+1)(d+2)

c1 ∆ v 2k ≤ vh 2L2 (τ ) ≤ c2 ∆ v 2k (9.17)
k=1 k=1
where vh is a quadratic function as defined in (9.16). Moreover, as for lin-

ear functions also the inverse inequality (9.15) remains valid for quadratic
functions.
Finally we will discuss bubble functions ϕB and their associated form
functions ψB which are needed, for example, for a stable discretization of the
Stokes problem. The basis functions ϕB are polynomial in the finite element
τ , and zero on the element boundary ∂τ . Hence we can extend them by zero
outside of the finite element τ . Later we will make use of an inverse inequality
for the induced trial space ShB (TN ) which is spanned by the bubble functions.
For d = 1 we have the form function
ψB (ξ) = ξ(1 − ξ) for ξ ∈ τ
and for the associated basis function ϕB

it follows that
1
2 1
L2 (τ ) = ∆
ϕB [ξ(1 − ξ)]2 dξ = h .
30
0
Moreover,
1 2
d 1
∇x ϕB 2
L2 (τ ) = h−1
[ξ(1 − ξ)] dξ = h−1 .
dξ 3
0
Hence we conclude for the one–dimensional bubble function the local inverse
inequality √
∇x ϕB L2 (τ ) = 10 h−1
ϕ L2 (τ ) .
B
In the two–dimensional case the form function ψB reads

ψB = ξ1 ξ2 (1 − ξ1 − ξ2 ) for ξ ∈ τ
and for the associated basis function ϕB it follows that

2 1
ϕB L2 (τ ) = 2∆ [ξ1 ξ2 (1 − ξ1 − ξ2 )]2 dξ = ∆ .
2520
τ
Then, by using Lemma 9.1 we conclude

ξ2 (1 − ξ2 − 2ξ1 ) 2 c
2 2
L2 (τ ) ≤ c ∇ξ ψB L2 (τ ) = c
∇x ϕB
ξ1 (1 − ξ1 − 2ξ2 ) dξ = 90
τ
and therefore we obtain the local inverse inequality
∇x ϕB c h−1
L2 (τ ) ≤ ϕ L2 (τ ) .
B
(9.18)
In the three–dimensional case d = 3 we finally have
ψ(ξ) = ξ1 ξ2 ξ3 (1 − ξ1 − ξ2 − ξ3 )
and therefore

1
v B L2 (τ ) = 6∆ [ψB (ξ)]2 dξ = ∆
415800
τ
as well as
c
∇x v B 2L2 (τ ) ≤ c ∆ h−2 2
∇ξ ψB L2 (τ ) = ∆ h−2
,
15120
in particular we conclude the local inverse inequality (9.18) also for d = 3.
9.3 Trial Spaces

The standard trial space to construct an approximate solution of boundary
value problems with second order partial differential equations is the space
Sh1 (TN ) of piecewise linear and globally continuous functions. When consider-
ing an admissible decomposition (9.1) those functions are uniquely determined
by the nodal function values vk = vh (xk ) which are given at the nodes xk of
the decomposition. Therefore, in the finite element τ we then have a local
representation by using local form functions. The dimension dim Sh1 (TN ) = M
of the global trial space Sh1 (TN ) is obviously equal to the number of nodes in
the decomposition. A basis of the trial space Sh1 (TN ) is given by, see Fig. 9.4,
⎧
⎪
⎨1 for x = xk ,
1
ϕk (x) := 0 for x = x = xk ,
⎪
⎩
linear elsewhere.
9.3 Trial Spaces 217
ϕ1k (x)
AB
BA
BA
BA
ϕ1k (x) BA
@
@ B A
@ B A
@ B A
@ @ A

@
AA
r r @r r @
@ xk
xk
@
@
Fig. 9.4. Linear basis functions for d = 1, 2.
If vh ∈ Sh1 (TN ) is piecewise linear, then we can write

M
vh (x) = vk ϕ1k (x).
k=1
Lemma 9.7. For vh ∈ Sh1 (TN ) there hold the spectral equivalence inequalities
⎛ ⎞ ⎛ ⎞
1
M 1 M
⎝ ∆ ⎠ vk2 ≤ vh 2L2 (TN ) ≤ ⎝ ∆ ⎠ vk2 .
(d + 1)(d + 2) d+1
k=1 ∈I(k) k=1 ∈I(k)
Proof. By using Lemma 9.4 we have with

⎛ ⎞

N
N
∆
d+1
1
M
vh 2L2 (TN ) = vh 2L2 (τ ) ≤ v 2k = ⎝ ∆ ⎠ vk2
d+1 d+1
=1 =1 k=1 k=1 ∈I(k)
the upper estimate. The lower estimate follows in the same way.

Lemma 9.8. For a piecewise linear function vh ∈ Sh1 (TN ) there holds the
inverse inequality

N
∇x vh 2L2 (TN ) ≤ cI h−2 2
vh L2 (τ ) .
=1
If the decomposition TN is globally quasi–uniform, then we have
∇x vh L2 (TN ) ≤ c h−1 vh L2 (TN ) . (9.19)
Proof. Both estimates follow immediately from Lemma 9.6.

To prove some approximation properties of the trial space Sh1 (TN ) we will
use error estimates of certain interpolation and projection operators. For a
globally continuous function v ∈ C(TN ) we define the interpolation in the
space of piecewise linear functions,

M
Ih v(x) := v(xk )ϕk (x) ∈ Sh1 (TN ). (9.20)
k=1
Lemma 9.9. Let v|τ ∈ H 2 (τ )be given. Then there holds the local error esti-
mate
v − Ih vL2 (τ ) ≤ c h2 |v|H 2 (τ ) .
Proof. For the error of the piecewise linear interpolation we first have, by
using the norm equivalence inequalities of Theorem 9.3,
v − Ih vL2 (τ ) ≤ c ∆
v − Iτ v L2 (τ ) ,
where Iτ is the linear interpolation operator with respect to the reference

element τ . Then,
Iτ v L2 (τ ) ≤ meas (τ )
v L∞ (τ )
and the use of the Sobolev imbedding theorem (Theorem 2.5) gives

v L∞ (τ ) ≤ c
v H 2 (τ ) .
Therefore we conclude that the linear operator
Iτ : H 2 (τ ) → L2 (τ )
is bounded. For an arbitrary but fixed w ∈ L2 (τ ) we define the linear func-

tional
f (u) := [(I − Iτ )u(ξ)]w(ξ)dξ .
τ
2
If u ∈ H (τ ) is given, then we have

|f (u)| = [(I − Iτ )u(ξ)]w(ξ)dξ

τ
≤ (I − Iτ )uL2 (τ ) wL2 (τ ) ≤ c uH 2 (τ ) wL2 (τ ) .
For any linear function q ∈ P1 (τ ) we have Iτ q = q and therefore f (q) = 0 for

all q ∈ P1 (τ ). Thus, all assumptions of the Bramble–Hilbert lemma (Theorem
2.8) are satisfied implying
|f (u)| ≤
c wL2 (τ ) |u|H 2 (τ ) .
When choosing u := v and w := (I − Iτ )

v we obtain

v 2L2 (τ ) =
(I − Iτ ) v (ξ)]2 dξ =
[(I − I ) [(I − I )
v ]w(ξ)dξ = |f (
v )|
τ τ
≤
c wL2 (τ ) |
v |H 2 (τ ) ≤
c (I − Iτ )
v L2 (τ ) |
v |H 2 (τ )
and hence the estimate
v L2 (τ ) ≤
(I − Iτ ) c |
v |H 2 (τ )
follows. Altogether we therefore have
v |H 2 (τ ) ≤ ĉ h2 |v|H 2 (τ )
v − Ih vL2 (τ ) ≤ c ∆ |
by applying the norm equivalence theorem (Theorem 9.3).

As a direct consequence of the above we conclude the global error estimate

N
v − Ih v2L2 (TN ) ≤ c h4 |v|2H 2 (τ ) . (9.21)
=1
In the same way we obtain also the error estimate

N
v − Ih v2H 1 (TN ) ≤ c h2 |v|2H 2 (τ ) . (9.22)
=1
The application of the interpolation operator requires the global continuity of

the given function to be interpolated. To weaken this strong assumption we
now consider projection operators which are defined via variational problems.
For a given u ∈ L2 (TN ) we define the L2 projection Qh u ∈ Sh1 (TN ) as the
unique solution of the variational problem
Qh u, vh L2 (TN ) = u, vh L2 (TN ) for all vh ∈ Sh1 (TN ). (9.23)
When choosing vh = Qh u as a test function we obtain the stability estimate
Qh uL2 (TN ) ≤ uL2 (TN ) for all u ∈ L2 (TN ), (9.24)
and by using the Galerkin orthogonality
u − Qh u, vh L2 (TN ) = 0 for all vh ∈ Sh1 (TN ) (9.25)
we conclude
u − Qh u2L2 (TN ) = u − Qh u, u − Qh uL2 (TN )

= u − Qh u, uL2 (TN )
≤ u − Qh uL2 (TN ) uL2 (TN )
and therefore
u − Qh uL2 (TN ) ≤ uL2 (TN ) for all u ∈ L2 (TN ). (9.26)
On the other hand, again by using the Galerkin orthogonality (9.25) we have
u − Qh u2L2 (TN ) = u − Qh u, u − Qh uL2 (TN )

= u − Qh u, u − Ih uL2 (TN )
≤ u − Qh uL2 (TN ) u − Ih uL2 (TN )
and therefore the error estimate

N
u − Qh u2L2 (TN ) ≤ u − Ih u2L2 (TN ) ≤ c h4 |v|2H 2 (τ ) (9.27)
=1
as well as
u − Qh uL2 (TN ) ≤ c h2 vH 2 (TN ) .
By interpolating this estimate with the error estimate (9.26) this yields the
error estimate
u − Qh uL2 (TN ) ≤ c h vH 1 (TN ) . (9.28)
By Q1h : H 1 (TN ) → Sh1 (TN ) we denote the H 1 projection which is defined as
Q1h u, vh H 1 (TN ) = u, vh H 1 (TN ) for all vh ∈ Sh1 (TN ). (9.29)
As above we find the stability estimate
Q1h uH 1 (TN ) ≤ uH 1 (TN ) for all u ∈ H 1 (TN ) (9.30)
and the error estimate
u − Q1h uH 1 (TN ) ≤ uH 1 (TN ) (9.31)
as well as

N
u − Q1h u2H 1 (TN ) ≤ u − Ih u2H 1 (TN ) ≤ c h2 |u|2H 2 (τ ) . (9.32)
=1
Hence we obtain the approximate property of the trial space Sh1 (TN ) of piece-
wise linear and continuous functions.
Theorem 9.10. Let u ∈ H s (TN ) with s ∈ [σ, 2] and σ = 0, 1. Then there

holds the approximation property
inf
1 (T )
u − vh H σ (TN ) ≤ c hs−σ |u|H s (TN ) . (9.33)
vh ∈Sh N
Proof. For σ = 0 and s = 2 the assertion is a direct consequence of the error

estimate (9.27). For s = 0 the approximation property is just the error esti-
mate (9.26). For s ∈ (0, 2) we then apply the interpolation theorem (Theorem
2.18). For σ = 1 we use the error estimates (9.32) and (9.31) to obtain the
result in the same way.
In what follows we will investigate further properties of the H 1 projection
Q1h which are needed later on.
Lemma 9.11. For s ∈ (0, 1] let w ∈ H01 (TN ) be the uniquely determined
w, vH 1 (TN ) = u − Q1h u, vH 1−s (TN ) for all v ∈ H 1 (TN ). (9.34)
If we assume w ∈ H 1+s (TN ) satisfying
wH 1+s (TN ) ≤ c u − Q1h uH 1−s (TN ) ,
then there holds the error estimate
u − Q1h uH 1−s (TN ) ≤ c hs u − Q1h uH 1 (TN ) . (9.35)
Proof. Using the assumptions we conclude
u − Q1h u2H 1−s (TN ) = u − Q1h u, u − Q1h uH 1−s (TN )
= w, u − Q1h uH 1 (TN )

= w − Q1h w, u − Q1h uH 1 (TN )
≤ w − Q1h wH 1 (TN ) u − Q1h uH 1 (TN )
≤ c hs |w|H 1+s (TN ) u − Q1h uH 1 (TN )
c hs u − Q1h uH 1−s (TN ) u − Q1h uH 1 (TN )
≤
from which the error estimate follows.

Remark 9.12. In Lemma 9.11, the best possible value of s ∈ (0, 1] depends on
the regularity of the decomposition TN . If, for example, TN is convex, then
we obtain s = 1 [66]. In the case of a corner domain, see, for example, [49].
Note that due to Sh1 (TN ) ⊂ H 1+s (TN ) for s ∈ (0, 12 ) the H 1 projection Q1h u is
well defined also for functions u ∈ H 1−s (TN ) and s ∈ (0, 12 ). As in the proof
of Lemma 9.11 we then can conclude the stability estimate
Q1h uH 1−s (TN ) ≤ c uH 1−s (TN ) for all u ∈ H 1−s (TN ). (9.36)
Using the error estimates of Lemma 9.11 for s = 1 we can show the stability
of the L2 projection in H 1 (TN ).
Lemma 9.13. Let the assumptions of Lemma 9.11 be satisfied for s = 1.

Then, the L2 projection Qh : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN ) is bounded, i.e.
Qh vH 1 (TN ) ≤ c vH 1 (TN ) for all v ∈ H 1 (TN ).
Proof. Let Q1h : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN ) be the H 1 projection as defined
in (9.29). By using the triangle inequality, the stability estimate (9.30), the
global inverse inequality (9.19), and the projection property Qh vh = vh for
all vh ∈ Sh1 (TN ) we obtain
Qh vH 1 (TN ) ≤ Q1h vH 1 (TN ) + Qh v − Q1h vH 1 (TN )

≤ vH 1 (TN ) + cI h−1 Qh v − Q1h vL2 (TN )
= vH 1 (TN ) + cI h−1 Qh (u − Q1h u)L2 (TN ) .
By applying the stability estimate (9.24) for Qh we further conclude
Qh vH 1 (TN ) ≤ vH 1 (TN ) + cI h−1 v − Q1h vL2 (TN ) .
Now the stability estimate follows from the error estimate (9.35) for Q1h and
from the stability estimate (9.30).
Remark 9.14. The L2 projection Qh : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN ) is also

bounded when the decomposition TN is locally adaptive refined, if the ratio
of local mesh sizes of neighboring elements does not vary too strongly [27].
In what follows we will always assume that the L2 projection is stable in

H 1 (TN ). Then, by using an interpolation argument, it follows that the error
estimate
u − Qh uH s (TN ) ≤ c h1−s uH 1 (TN ) for all u ∈ H 1 (TN ) (9.37)
is valid.
Based on the trial space of piecewise linear and globally continuous func-
tions we can introduce trial spaces of locally higher polynomial degrees.
In the one–dimensional case d = 1 we can define quadratic basis functions
locally by
ϕ2 (x) = 4ξ(1 − ξ) for x = x 1 + ξ h ∈ τ .
An arbitrary function vh ∈ Sh2 (TN ) then can be written as

M
N
vh (x) = vk ϕ1k (x) + vM + ϕ2 (x).
k=1 =1
Therefore we have dim Sh2 (TN ) = M + N . For both the two–dimensional case
d = 2 and the three–dimensional case d = 3 we have to ensure the continuity
of the quadratic basis functions ϕ2 . Since the quadratic form functions are
defined locally with respect to the edges of the reference element, the support
of a global quadratic basis function consists of those finite elements which
share the corresponding edge. Denote by K the number of all edges of the
decomposition (9.1), then we can write for d = 2, 3 the global representation

M
K
vh (x) = vk ϕ1k (x) + vM +j ϕ2j (x)
k=1 j=1
as well as the local representation

1
2 (d+1)(d+2)

vh (x) = v i ψi2 (ξ) for x = x 1 + J ξ, ξ ∈ τ.
i=1
Here, 1 , . . . , d+1 denote, as before, the indices of the associated global nodes,
while d+2 , . . . , 12 (d+1)(d+2) are the indices of the associated global edges, see
also Fig. 9.3 for d = 2.
Lemma 9.15. For vh ∈ Sh2 (TN ) there hold the spectral equivalence inequalities
2 2
dimS
h (TN ) dimS
h (TN )
c1 dk vk2 ≤ vh 2L2 (TN ) ≤ c2 dk vk2

k=1 k=1
with ⎧ ;
⎨ ∆ for k = 1, . . . , M,
∈I(k)
dk :=
⎩
∆k−M for k = M + 1, . . . , M + N
in the one–dimensional case d = 1, and
⎧ ;
⎪
⎪ ∆ for k = 1, . . . , M,
⎨ ∈I(k)
dk := ;
⎪
⎪ ∆ for k = M + 1, . . . , M + K
⎩
∈K(k−M )
when d = 2, 3.
Proof. First we will use the local spectral equivalence inequalities (9.17). For
d = 1 we have1

N
N 3

vh 2L2 (TN ) = vh 2L2 (τ ) ∆ v 2i
=1 =1 i=1
⎛ ⎞

M
N
= ⎝ ∆ ⎠ vk2 + 2
∆ vM + .
k=1 ∈I(k) =1
1
The equivalence A B means that there are positive constants c1 and c2 such
that c1 A ≤ B ≤ c2 A.
In the same way we have for d = 2, 3

N
N 3

vh 2L2 (τ ) ∆ v 2k
=1 =1 k=1
⎛ ⎞ ⎛ ⎞

M
K
= ⎝ ∆ ⎠ vk2 + ⎝ ∆ ⎠ vM
2
+j .

k=1 ∈I(k) j=1 ∈K(j)
By Ih : C(TN ) → Sh2 (TN ) we denote the interpolation operator into the

trial space of locally quadratic functions. The interpolation nodes are hereby
all M nodes of the decomposition (9.1) and all N element midpoints in the
one–dimensional case d = 1 and all K edge midpoints in the cases d = 2, 3.
Note that the interpolation operator Ih is exact for locally quadratic functions.
Analogous to Lemma 9.9 as well as to the global error estimates (9.21) and
(9.22) we can prove the following error estimates, when assuming u ∈ H 3 (TN ),

N
u − Ih uL2 (TN ) ≤ c h6 |u|2H 3 (τ ) ,
=1
and

N
u − Ih u H 1 (TN ) ≤ c h4 |u|2H 3 (τ ) .
=1
As in the case of piecewise linear basis functions we can show a global ap-
proximation property.
Theorem 9.16. Let u ∈ H s (TN ) with s ∈ [σ, 3] and σ = 0, 1. Then there
holds
inf
2
u − vh H σ (TN ) ≤ c hs−σ |u|H s (TN ) .
vh ∈Sh (TN )
By ShB (TN ) = span{ϕB } =1 we denote the global trial space of local bubble
N
functions. For an arbitrary given vh ∈ ShB (TN ) we can write

N
vh (x) = v B ϕB
(x).
=1
If the decomposition TN is globally quasi–uniform we can derive, by using the

local inverse inequality (9.18), the global inverse inequality
∇vh L2 (TN ) ≤ cI h−1 vh L2 (TN ) for all vh ∈ ShB (TN ). (9.38)
For a given u ∈ L2 (TN ) we denote by : L2 (TN ) → QB
h the projection ShB (TN )
into the trial space ShB (TN ) which is the unique solution of the variational
problem
QB
h v(x)dx = v(x)dx for all = 1, . . . , N. (9.39)
τ τ
9.4 Quasi Interpolation Operators 225
Lemma 9.17. For v ∈ L2 (TN ) let QB h v ∈ Sh (TN ) be the projection as defined

B
in (9.39). Then there holds the stability estimate

√
QBh vL2 (TN ) ≤ 2 vL2 (TN ) .
Proof. From (9.39) we find for the coefficients of vh ∈ ShB (TN )

⎧
⎨ 6 for d = 1,
cd
v = v(x)dx, cd = 60 for d = 2,
∆ ⎩
τ 840 for d = 3.
Hence we have
⎡ ⎤2

c2d ∆ ⎣
QB 2 B 2
h vL2 (τ ) = |v | ϕ L2 (τ ) = v(x)dx⎦
∆2 cB
d
τ
with ⎧
⎨ 30 for d = 1,
cB
d = 2520 for d = 2,
⎩
415800 for d = 3.
By using c2d /cB
d < 2 for d = 1, 2, 3 and by applying the Cauchy–Schwarz
inequality we therefore obtain
⎡ ⎤2

2 ⎣ ⎦ 2
QB
h vL2 (τ ) ≤ v(x)dx ≤ dx [v(x)]2 dx = 2 v2L2 (τ ) .
∆ ∆
τ τ τ
Taking the sum over all elements this gives the desired stability estimate.

9.4 Quasi Interpolation Operators

For a given function v ∈ H 1 (TN ) we have considered the piecewise linear
interpolation (9.20). By using Lemma 9.9 there holds for v ∈ H 2 (TN ) the
local error estimate
v − Ih vL2 (τ ) ≤ c h2 |v|H 2 (τ )
where we have to assume the continuity of the function to be interpolated.

In particular, the interpolation operator Ih is not defined for a general
v ∈ H 1 (TN ) and d = 2, 3, and therefore Ih is not a continuous operator
in H 1 (TN ).
On the other hand, the L2 projection
Qh : L2 (TN ) → Sh1 (TN ) ⊂ L2 (TN )

as defined in (9.23) is bounded (see (9.24)), and there holds the global error
estimate (9.27),
u − Qh uL2 (TN ) ≤ c h2 |u|H 2 (TN ) .
As already stated in Remark 9.14, the L2 projection
Qh : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN )
is bounded, but it is not possible to derive a local error estimate. Hence we

aim to construct a bounded projection operator
Ph : H 1 (TN ) → Sh1 (TN ) ⊂ H 1 (TN ),
which admits a local error estimate. This can be done by using quasi interpo-
lation operators [42].
For any node xk of the locally uniform decomposition TN we define
*
ω k := τ
∈I(k)
to be the convex support of the associated piecewise linear basis function

ϕ1k ∈ Sh1 (TN ). By ĥk we denote the averaged mesh size of ωk which is equiv-
alent to the local mesh sizes h of all finite elements τ with ∈ I(k) when
the decomposition is assumed to be locally quasi–uniform. Then we introduce
Qkh : L2 (ωk ) → Sh1 (ωk ) as the local L2 projection which is defined by the
variational formulation
Qkh u, vh L2 (ωk ) = u, vh L2 (ωk ) for all vh ∈ Sh1 (ωk ),
and by using (9.27) there holds the error estimate
u − Qkh uL2 (ωk ) ≤ c ĥk |u|H 1 (ωk )
As in (9.24) we can prove the stability estimate
Qkh uL2 (ωk ) ≤ uL2 (ωk ) for all u ∈ L2 (ωk )
and by applying Lemma 9.13 there holds
Qkh uH 1 (ωk ) ≤ c uH 1 (ωk ) for all u ∈ H 1 (ωk ).
By using the local projection operators we can define the quasi interpolation
operator or Clement operator

M
(Ph u)(x) = (Qkh u)(xk ) ϕ1k (x).
k=1
It is easy to check that Ph vh = vh ∈ Sh1 (TN ).

9.5 Exercises 227
Theorem 9.18. [27, 42, 139] For u ∈ H 1 (TN ) there holds the local error
estimate

u − Ph uL2 (τ ) ≤ c ĥk |u|H 1 (ωk ) for all = 1, . . . , N, (9.40)
k∈J( )
and the global stability estimate
Ph uH 1 (TN ) ≤ c uH 1 (TN ) . (9.41)
Proof. Let τ be an arbitrary but fixed finite element and let k ∈ J() be an
arbitrary fixed index. For x ∈ τ we can write

(Ph )(x) = (Qkh u)(x) + [(Qkh u)(xk ) − (Qkh )(xk )]ϕ1k (x).

k∈J( ),k=k
By using Lemma 9.4 we have

∆
ϕk L2 (τ ) ≤ ,
d+1
and therefore
d/2

u−Ph uL2 (τ ) ≤ c1 ĥk̃ |u|H 1 (ωk̃ ) +c2 h |(Qkh u)(xk )−(Qk̃h u)(xk )| .
k∈J( ),k=k̃
For an arbitrary vh ∈ Sh1 (TN ) we conclude from Corollary 9.5

−d/2
vh L∞ (τ ) ≤ c h vh L2 (τ ) .
Therefore,
|(Qkh u)(xk ) − (Qk̃h u)(xk )| ≤ Qkh u − Qk̃h uL∞ (τ )

−d/2
≤ c h Qkh u − Qk̃h uL2 (τ )
! "
−d/2
≤ c h Qkh u − uL2 (τ ) + u − Qk̃h L2 (τ )
−d/2
≤ c h hk |u|H 1 (ωk ) + hk̃ |u|H 1 (ωk̃ ) ,
from which the error estimate (9.40) follows. The stability estimate (9.41) can
be shown in the same way.
9.5 Exercises
9.1 For an admissible decomposition of a bounded domain Ω ⊂ R2 into
triangular finite elements τ and for piecewise linear continuous basis functions
ϕk the mass matrix Mh is defined by

Mh [j, k] = ϕk (x)ϕj (x)dx, j, k = 1, . . . , M.
Ω
Find a diagonal matrix Dh and positive constants c1 and c2 such that the
spectral equivalence inequalities
c1 (Dh u, u) ≤ (Mh u, u) ≤ c2 (Dh u, u)
are satisfied for all u ∈ RM .

9.2 For the two–dimensional reference element τ ⊂ R2 the local quadratic
shape functions are given by
ψ12 (ξ) = 1 − ξ1 − ξ2 , ψ22 (ξ) = ξ1 , ψ32 (ξ) = ξ2 ,

ψ42 (ξ) = 4ξ1 (1 − ξ1 − ξ2 ), ψ52 (ξ) = 4ξ1 ξ2 , ψ62 (ξ) = 4ξ2 (1 − ξ1 − ξ2 ).
Compute the local mass matrix M as well as the minimal and maximal
eigenvalues of M .
10
Boundary Elements
For the approximate solution of the boundary integral equations as considered

in Chapter 7 we introduce suitable finite–dimensional trial spaces. These are
based on appropriate parametrizations of the boundary Γ = ∂Ω and on the
use of finite elements in the parameter domain. In particular we can think of
boundary elements as finite elements on the boundary.
10.1 Reference Elements

<J
Let Γ = ∂Ω be a piecewise smooth Lipschitz boundary with Γ = j=1 Γ j
where any boundary part Γj allows a local parametrization Γj = χj (Q) with
respect to some parameter domain Q ⊂ Rd−1 . We assume that
cχ1 ≤ |detχj (ξ)| ≤ cχ2 for all ξ ∈ Q, j = 1, . . . , J. (10.1)
Further we consider a sequence {ΓN }N ∈N of decompositions (meshes)
*
N
ΓN = τ (10.2)
=1
with boundary elements τ . We assume that for each boundary element τ

there exists a unique index j with τ ⊂ Γj . A decomposition of the boundary
part Γj into boundary elements τ implies a decomposition of the parameter
domain Q into finite elements q j with τ = χj (q j ). In the simplest case the
boundary elements τ are intervals in the two–dimensional case d = 2 or
triangles in the three–dimensional case d = 3, see Fig. 10.1.
Example 10.1. The boundary of the two–dimensional L shaped domain as de-

picted in Fig. 10.1 can be described by using the following parametrization
for ξ ∈ Q = (0, 1):
230 10 Boundary Elements
r r r r r r r r r
r r
r r
r r
r r r r r r
r r
r r
r r
r r r r r
Fig. 10.1. Boundary discretization with 32 and 56 boundary elements.
⎛⎞ ⎛1⎞ ⎛ 1 − 2ξ ⎞
ξ
⎜4⎟ ⎜ ⎟
χ1 (ξ) = ⎝ 4 ⎠ , χ2 (ξ) = ⎝ ⎠ , χ3 (ξ) = ⎝ 4 ⎠ ,
ξ 1
0
4 4
⎛ 1 ⎞ ⎛ξ−1⎞ ⎛ ⎞
− 0
⎜ 4 ⎟ ⎜ 4 ⎟
χ4 (ξ) = ⎝ ⎠ , χ5 (ξ) = ⎝ ⎠ , χ6 (ξ) = ⎝ ξ − 1 ⎠ .
1 − 2ξ 1
− 4
4 4
For j = 1, 2, 5, 6 the parameter domain Q = (0, 1) is decomposed into 4 equal
sized elements q j while for j = 3, 4 we have 8 elements q j to be used.
By {xk }Mk=1 we denote the set of all nodes of the boundary decomposition
ΓN . The index set I(k) describes all boundary elements τ where xk is a node,
while J() is the index set of all nodes xk describing the boundary element
τ . In the three–dimensional case d = 3 the boundary decomposition (10.2) is
called admissible, if two neighboring boundary elements share either a node
or an edge, see also Fig. 9.2. Analogous to (9.2) we compute by

∆ := dsx
τ
the volume and by

1/(d−1)
h := ∆
the local mesh size of the boundary element τ . Then,
h := max h
=1,...,N
is the global mesh size of the boundary decomposition (10.2). Moreover,
d := sup |x − y|
x,y∈τ
is the diameter of the boundary element τ . Finally,
hmin := min h .
=1,...,N
is the minimal mesh size. The family of boundary decompositions (10.2) is

called globally quasi–uniform if
hmax
≤ cG
hmin
is satisfied with a global constant cG ≥ 1 which is independent of N ∈ N. The
family {ΓN }N ∈N is called locally quasi–uniform if
h
≤ cL for = 1, . . . , N
hj
is satisfied for all neighboring elements τj of τ , i.e. τ and τj share either a

node or an edge.
In the two–dimensional case d = 2 a boundary element τ with nodes x 1
and x 2 can be described via the parametrization
x(ξ) = x 1 + ξ(x 2 − x 1 ) for ξ ∈ τ = (0, 1)
where τ = (0, 1) is the reference element, and we have
1 /
d = h = ∆ = dsx = [x1 (ξ)]2 + [x2 (ξ)]2 dξ = |x 2 − x 1 | .
τ 0
In the three–dimensional case d = 3 we consider plane triangular boundary

elements τ with nodes x 1 , x 2 and x 3 . The parametrization of τ with respect
to the reference element

τ := ξ ∈ R2 : 0 < ξ1 < 1, 0 < ξ2 < 1 − ξ1
then reads
x(ξ) = x 1 + ξ1 (x 2 − x 1 ) + ξ2 (x 3 − x 1 ) for ξ ∈ τ.
For the computation of the boundary element volume we obtain

.
1.
∆ = dsx = EG − F 2 dξ = EG − F 2
2
τ τ
where
3 2
∂
E= xi (ξ) = |x 2 − x 1 |2 ,
i=1
∂ξ 1
3 2
∂
G= xi (ξ) = |x 3 − x 1 |2 ,
i=1
∂ξ 2
3
∂ ∂
F = xi (ξ) xi (ξ) = (x 2 − x 1 , x 3 − x 1 ) .
i=1
∂ξ1 ∂ξ2
In the three–dimensional case d = 3 we assume that all boundary elements τ

are shape regular, i.e. there exists a constant cB independent of the boundary
decomposition such that
d ≤ cB h for = 1, . . . , N. (10.3)
By using ⎛ ⎞
x 2 ,1 − x 1 ,1 x 3 ,1 − x 1 ,1
J = ⎝ x 2 ,2 − x 1 ,2 x 3 ,2 − x 1 ,2 ⎠
x 2 ,3 − x 1 ,3 x 3 ,3 − x 1 ,3
we can write a function v(x) for x ∈ τ as
v(x) = v(x 1 + J ξ) =: v (ξ) for ξ ∈ τ.
Vice versa, for a function v(ξ) which is given for ξ ∈ τ we can define a function
v (x) for x ∈ τ ,
v (x) := v(x 1 + J ξ) = v(ξ) for ξ ∈ τ.
In the two–dimensional case d = 2 we have

v2L2 (τ ) = |v(x)|2 dsx = v (ξ)|2 h dξ = ∆
| v 2L2 (τ )
τ τ
and for the three–dimensional case d = 3 it follows that

v2L2 (τ ) = |v(x)|2 dsx = 2∆ |
v (ξ)|2 dξ = 2∆
v 2L2 (τ ) .
τ τ
To define Sobolev spaces H (Γ ) for s ≥ 1 we have to use a parametrization

s
Γj = χj (Q), see Section 2.5. In particular,

|v|2H 1 (τ ) := |∇ξ v(χj (ξ))|2 dξ
qj
and
∆ = dsx = det|χj (ξ)| dξ.
τ qj
10.2 Trial Spaces
With respect to the boundary decomposition (10.2) we now define trial spaces
of local polynomials. In particular we will consider the trial space Sh0 (Γ ) of
piecewise constant functions and the trial space Sh1 (Γ ) of piecewise linear
continuous functions. By considering appropriate interpolation and projection
operators we will prove certain approximation properties of these trial spaces.
Let
Sh0 (Γ ) := span{ϕ0k }N
k=1
be the space of functions which are piecewise constant with respect to the
boundary decomposition (10.2). The basis functions ϕ0k are given by

1 for x ∈ τk ,
ϕ0k (x) =
0 elsewhere.
If u ∈ L2 (Γ ) is a given function, the L2 projection Qh u ∈ Sh0 (Γ ) is defined as

Qh u, vh L2 (Γ ) = u, vh L2 (Γ ) for all vh ∈ Sh0 (Γ ). (10.4)
This is equivalent to finding the coefficient vector u ∈ RN as the solution of

N
uk ϕ0k , ϕ0 L2 (Γ ) = u, ϕ0 L2 (Γ ) for = 1, . . . , N.
k=1
Due to

∆k for k = ,
ϕ0k , ϕ0 L2 (Γ ) = ϕ0k (x)ϕ0 (x)dsx =
0 for k =
Γ
we obtain
1
uk = u(x)dsx for k = 1, . . . , N.
∆k
τk
Theorem 10.2. Let u ∈ H s (Γ ) be given for some s ∈ [0, 1], and let
Qh u ∈ Sh0 (Γ ) be the L2 projection as defined by (10.4). Then there hold the
error estimates

N
u − Qh u2L2 (Γ ) ≤ c h2s 2
k |u|H s (τk ) (10.5)
k=1
and
u − Qh uL2 (Γ ) ≤ c hs |u|H s (Γ ) . (10.6)
Proof. By using the Galerkin orthogonality
u − Qh u, vh L2 (Γ ) = 0 for all vh ∈ Sh0 (Γ )
we obtain
u − Qh u2L2 (Γ ) = u − Qh u, u − Qh uL2 (Γ )
= u − Qh u, uL2 (Γ ) ≤ u − Qh uL2 (Γ ) uL2 (Γ )
and therefore
u − Qh uL2 (Γ ) ≤ uL2 (Γ )
which is the error estimate for s = 0.
We now consider s ∈ (0, 1). For x ∈ τk we have Qh u(x) = uk and therefore

1
u(x) − Qh u(x) = [u(x) − u(y)]dsy for x ∈ τk .
∆k
τk
By taking the square and applying the Cauchy–Schwarz inequality we con-

clude
⎛ ⎞2

1
|u(x) − Qh u(x)|2 = 2 ⎝ [u(x) − u(y)]dsy ⎠
∆k
τk
⎛ ⎞2

1 [u(x) − u(y)]
= 2⎝ 2 +s dsy ⎠
d−1
|x − y|
2 +s
d−1
∆k |x − y|
τk

1 [u(x) − u(y)]2
≤ 2 dsy |x − y|d−1+2s dsy
∆k |x − y|d−1+2s
τk τk

1 |u(x) − u(y)|2
≤ dd−1+2s dsy .
k
∆k |x − y|d−1+2s
τk
By using the shape regularity (10.3) and ∆k = hd−1 k we can replace the
diameter dk and the area ∆k by the local mesh size hk ,

[u(x) − u(y)]2
|u(x) − Qh u(x)|2 ≤ cd−1+2s h 2s
dsy .
B k
|x − y|d−1+2s
τk
When integrating with respect to x ∈ τk this gives
u − Qh u2L2 (τk ) ≤ cd−1+2s

B h2s 2
k |u|H s (τk ) ,
and by taking the sum over all boundary elements we obtain the error estimate
for s ∈ (0, 1).
To prove the error estimate for s = 1 we first consider

1
u(x) − Qh u(x) = [u(x) − u(y)]dsy for x ∈ τk .
∆k
τk
By using the local parametrization τk = χj (qkj ) we further get

1
u(x) − Qh u(x) = [u(χj (ξ)) − u(χj (η))]|detχj (η)| dη. (10.7)
∆k
j
qk
In the two–dimensional case d = 2 we have

ξ
1 d
u(x) − Qh u(x) = u(χj (t))dt |detχj (η)| dη
∆k dt
j η
qk
and therefore

1 d
|u(x) − Qh u(x)| ≤ u(χj (t))|dt detχj (η)| dη
∆k dt
j j
qk qk

1
= |detχj (η)| dη |∇ξ u(χj (ξ))|dξ
∆k
j j
qk qk

= |∇ξ u(χj (ξ))|dξ.
j
qk
By taking the square and applying the Cauchy–Schwarz inequality we find by

considering (10.1)
2

2
|u(x) − Qh u(x)| = |∇ξ u(χj (ξ))|dξ ≤ dξ |∇ξ u(χj (ξ))|2 dξ

qj qkj j
qk

k
1 1
≤ χ |detχj (ξ)|dξ |u|2H 1 (τk ) = χ ∆k |u|2H 1 (τk ) .
c1 c1
j
qk
When integrating with respect to x ∈ τk and using ∆k = hk for d = 2 this

gives
1
[u(x) − Qh u(x)]2 dsx ≤ χ h2k |u|2H 1 (τk ) ,
c1
τk
and by taking the sum over all boundary elements τk we finally obtain the
error estimate for s = 1 and d = 2.
In the three–dimensional case d = 3 we have from the representation (10.7)

1
u(x) − Qh u(x) = [u(χj (ξ)) − u(χj (η))]|detχj (η)| dη
∆k
j
qk
1
1 d
= u(χj (η + t(ξ − η)))dt |detχj (η)| dη
∆k dt
j 0
qk
1
1
= (ξ − η) · ∇η u(χj (η + t(ξ − η)))dt |detχj (η)| dη.
∆k
j
qk 0
Taking the square

⎛ ⎞2
1
1 ⎜ ⎟
|u(x)−Qh u(x)|2 ≤ 2 ⎝ (ξ − η) · ∇η u(χj (η + t(ξ − η)))dt |detχj (η)| dη⎠
∆k
j 0
qk
and by applying the Cauchy–Schwarz inequality this gives

2
1
1
|u(x) − Qh u(x)|2 ≤ 2 ∇η u(χj (η + t(ξ − η)))dt dη
∆k

j
qk 0

|ξ − η|2 |detχj (η)|2 dη
j
qk
2
1

≤ c ∇η u(χj (η + t(ξ − η)))dt dη.

j
qk 0
When integrating over τk we obtain

2
1

2
[u(x) − Qh u(x)] dsx ≤ c
∇η u(χj (η + t(ξ − η)))dt dη dξ

τk j j
qk qk 0
1
2
≤c |∇η u(χj (η + t(ξ − η)))| dt dη dξ
j j 0
qk qk

2
≤ c ∆k |∇η u(χj (η)| dξ = c ∆k |u|2H 1 (τk )
j
qk
and with ∆k = h2k for d = 3 we find
u − Qh u2L2 (τk ) ≤ c h2k |u|2H 1 (τk ) .
By taking the sum over all boundary elements we finally get the error estimate
for s = 1 and d = 3,

Corollary 10.3. Let u ∈ H s (Γ ) be given for some s ∈ [0, 1]. For σ ∈ [−1, 0)
then there hold the error estimates

N
u − Qh u2H σ (Γ ) ≤ c h−2σ h2s 2
k |u|H s (τk )
k=1
and
u − Qh uH σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.8)
Proof. For σ ∈ [−1, 0) we have by duality, by using the definition (10.4) of

the L2 projection, and by applying the Cauchy–Schwarz inequality
|u − Qh u, vL2 (Γ ) |
u − Qh uH σ (Γ ) = sup
0=v∈H −σ(Γ ) vH −σ (Γ )
|u − Qh u, v − Qh vL2 (Γ ) |
= sup
0=v∈H −σ(Γ ) vH −σ (Γ )
v − Qh vL2 (Γ )
≤ u − Qh uL2 (Γ ) sup .
0=v∈H −σ(Γ ) vH −σ (Γ )
By using the error estimate (10.5) for u − Qh uL2 (Γ ) and the estimate (10.6)
for v − Qh vL2 (Γ ) the assertion follows.

Altogether we can formulate the approximation property of the trial space
Sh0 (Γ ) of piecewise constant functions.
Theorem 10.4. Let σ ∈ [−1, 0]. For u ∈ H s (Γ ) with some s ∈ [σ, 1] there
holds the approximation property of Sh0 (Γ )
inf
0 (Γ )
u − vh H σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.9)
vh ∈Sh
Proof. For σ ∈ [−1, 0] and s ∈ [0, 1] the approximation property is just the
statement of Theorem 10.2 and Corollary 10.3. It remains to prove the ap-
proximation property for σ ∈ [−1, 0) and s ∈ [σ, 0).
For a given u ∈ H σ (Γ ) let Qσh u ∈ Sh0 (Γ ) ⊂ H σ (Γ ) ⊂ L2 (Γ ) be the H σ (Γ )
projection which is defined as the unique solution of the variational problem
Qσh u, vh H σ (Γ ) = u, vh H σ (Γ ) for all vh ∈ Sh0 (Γ ).
As for the L2 projection there holds the error estimate for s = σ,

u − Qσh uH σ (Γ ) ≤ uH σ (Γ ) .
Therefore, I − Qσh : H σ (Γ ) → H σ (Γ ) is a bounded operator with norm
I − Qσh H σ (Γ )→H σ (Γ ) ≤ 1.
On the other hand, by using (10.8) and s = 0 we have
u − Qσh uH σ (Γ ) ≤ u − Qh uH σ (Γ ) ≤ c h−σ uL2 (Γ ) .
Thus, I − Qσh : L2 (Γ ) → H σ (Γ ) is bounded with norm
I − Qσh L2 (Γ )→H σ (Γ ) ≤ c h−σ .
By applying the interpolation theorem (Theorem 2.18 and Remark 2.23) we

conclude that the operator I − Qσh : H s (Γ ) → H σ (Γ ) is bounded for all
s ∈ [σ, 0], and for the related operator norm it follows that
I − Qσh H s (Γ )→H σ (Γ )
s−0 s−σ
≤ I − Qσh H σ (Γ )→H σ (Γ ) σ−0 I − Qσh L2 (Γ )→H σ (Γ ) −σ
s−σ
≤ (c h−σ ) −σ = c(s, σ) hs−σ .
This gives the approximation property for σ ∈ [−1, 0) and s ∈ [σ, 0).
Now we consider the case where Γj ⊂ Γ is an open boundary part of
Γ = ∂Ω, and Sh0 (Γj ) is the associated trial space of piecewise constant basis
functions. As in (10.6) there holds the error estimate
u − Qh uL2 (Γj ) ≤ c hs |u|H s (Γj )
for the L2 projection Qh : L2 (Γj ) → Sh0 (Γj ) which is defined accordingly.

Analogous to Corollary 10.3 for σ ∈ [−1, 0) we find the error estimate
-σ (Γj ) ≤ c h
u − Qh uH s−σ
|u|H s (Γj ) .
Hence we have the approximation property
inf
0 (Γ )
u − vh H
-σ (Γj ) ≤ c h
s−τ
|u|H s (Γj ) (10.10)
vh ∈Sh j
for u ∈ H s (Γj ) and −1 ≤ σ ≤ 0 ≤ s ≤ 1.

In addition to the trial space Sh0 (Γ ) of piecewise constant basis functions
ϕk we next consider the trial space Sh1 (Γ ) of piecewise linear and globally
0
continuous basis functions ϕ1i . If the boundary decomposition (10.2) is ad-

missible, a function vh ∈ Sh1 (Γ ) is determined by the nodal values which are
described at the M nodes xk . Hence, a basis of Sh1 (Γ ) is given by
⎧
⎨1 for x = xi ,
ϕ1i (x) = 0 for x = xj = xi ,
⎩
linear elsewhere.
If a piecewise linear function vh is considered in the boundary element τ , this

function is uniquely determined by the nodal values vh (xk ) for k ∈ J(). By
using the parametrization τ = χj (q j ) we can write
vh (x) = vh (χj (ξ)) = v j (ξ) for ξ ∈ q j ⊂ Q ⊂ Rd−1 .
Hence we can identify a boundary element τ ⊂ Γ where Γ = ∂Ω and Ω ⊂ Rd

with a finite element q j in the parameter domain Q ⊂ Rd−1 . Thus we can
transfer all local error estimates of piecewise linear basis functions, which
were already proved in Chapter 9, to the finite element q j and therefore to
the boundary element τ .
Lemma 10.5. For a function vh which is linear on τ there holds
∆
d
∆ 2
d
v 2k ≤ vh 2L2 (τ ) ≤ v k .
d(d + 1) d
k=1 k=1
Proof. By mapping the boundary element τ to the reference element τ we

obtain
vh 2L2 (τ ) = vh , vh L2 (τ )

d d
= vi vj ϕ1i (ξ)ϕ1j (ξ)|detJ |dξ = (G v , v )
i=1 j=1 τ
where
∆
G = (Id + ed e
d)
d(d + 1)
is the local mass matrix and ed = 1 ∈ Rd . The eigenvalues of the matrix
Id + ed e
d are given by
λ1 = d + 1, λ2 = · · · = λd = 1
and therefore the assertion follows.

Corollary 10.6. For a function vh which is linear on τ there holds

∆
vh 2L∞ (τ ) ≤ vh 2L2 (τ ) ≤ ∆ vh 2L∞ (τ ) .
d(d + 1)
Proof. Since the maximum of |vh | and therefore the vh L∞ (τ ) norm is equal
to some nodal value |vh (xk∗ )| = |vk∗ | for some xk∗ , the assertion follows
immediately from Lemma 10.5.
Lemma 10.7. For a function vh which is linear on τ there holds the local
inverse inequality
|vh |H 1 (τ ) ≤ cI h−1
vh L2 (τ ) .
Proof. First we have

2
|vh |H 1 (τ ) = |∇ξ vh (χj (ξ))|2 |detχj (ξ)|dξ ≤ ∆ ∇ξ vh (χj (·))2L j .
∞ (q )
qj
By mapping the finite element q j to the associated reference element we obtain
∇ξ vh (χj (·))2L j ≤ c h−2 2

vh (χj (·))L j = c h−2 2
vh L∞ (τ ) ,
∞ (q ) ∞ (q )
and the inverse inequality follows from Corollary 10.6.

Hence we also conclude the global inverse inequality

N
N
|vh |2H 1 (Γ ) = |vh |2H 1 (τ ) ≤ c h−2 2
vh L2 (τ )
=1 =1
and, for a globally quasi–uniform boundary decomposition,
|vh |H 1 (Γ ) ≤ c h−1 vh L2 (Γ ) .
In particular for h < 1 we obtain
vh H 1 (Γ ) ≤ c h−1 vh L2 (Γ ) ,
and an interpolation argument gives
vh H s (Γ ) ≤ c h−s vh L2 (Γ ) for s ∈ [0, 1].
Analogous to the error estimates (9.21) and (9.22) we can estimate the interpo-
lation error of the piecewise linear interpolation operator Ih : H 2 (Γ ) → Sh1 (Γ )
as follows.
Lemma 10.8. Let v ∈ H 2 (Γ ) be given. Assume that Γ = ∂Ω is sufficiently
smooth where Ω ⊂ Rd . Let Ih v be the piecewise linear interpolation satisfying
Ih v(xk ) = v(xk ) at all nodes xk of the admissible boundary decomposition
(10.2). Then there hold the error estimates

N
v − Ih v2L2 (Γ ) ≤ c h4 |v|2H 2 (τ ) ≤ c h4 |v|2H 2 (Γ )
=1
and

N
v − Ih v2H 1 (Γ ) ≤ c h2 |v|2H 2 (τ ) ≤ c h2 |v|2H 2 (Γ ) .
=1
By applying the interpolation theorem (Theorem 2.18, Remark 2.23) we can

conclude the error estimate
v − Ih vH σ (Γ ) ≤ c h2−σ |v|H 2 (Γ ) for σ ∈ [0, 1].

The piecewise linear interpolation requires, as in the case of finite elements, the
global continuity of the function to be interpolated. The function v ∈ H s (Γ )
is continuous for s ∈ ( d−1
2 , 2] and the following error estimate holds,
v − Ih vH σ (Γ ) ≤ c hs−σ |v|H s (Γ ) , 0 ≤ σ ≤ min{1, s}. (10.11)
To prove more general error estimates we now consider projection operators

which are defined by some variational problems. If u ∈ L2 (Γ ) is given the L2
projection Qh u ∈ Sh1 (Γ ) is defined as the unique solution of the variational
problem
Qh u, vh L2 (Γ ) = u, vh L2 (Γ ) for all vh ∈ Sh1 (Γ ),
and there holds the error estimate
u − Qh uL2 (Γ ) ≤ uL2 (Γ ) .
On the other hand, by using Lemma 10.8 we also have the error estimate

N
u − Qh uL2 (Γ ) ≤ u − Ih uL2 (Γ ) ≤ c h4 |u|2H 2 (Γ ) ≤ c h2 |u|2H 2 (Γ )
=1
and by applying the interpolation theorem (Theorem 2.18, Remark 2.23) we

u − Qh uL2 (Γ ) ≤ c hs |u|H s (Γ ) for u ∈ H s (Γ ), s ∈ [0, 2]. (10.12)
Accordingly, for u ∈ H σ (Γ ) and σ ∈ (0, 1] we define the H σ projection

Qσh : H σ (Γ ) → Sh1 (Γ ) as the unique solution of the variational problem
Qσh u, vh H σ (Γ ) = u, vh H σ (Γ ) for all vh ∈ Sh1 (Γ )
satisfying the error estimate
u − Qσh uH σ (Γ ) ≤ c hs−σ |u|H s (Γ ) for u ∈ H s (Γ ), s ∈ [σ, 2]. (10.13)
Theorem 10.9. Let Γ = ∂Ω be sufficiently smooth. For σ ∈ [0, 1] and for

some s ∈ [σ, 2] we assume u ∈ H s (Γ ). Then there holds the approximation
property of Sh1 (Γ ),
inf
1 (Γ )
u − vh H σ (Γ ) ≤ c hs−σ |u|H s (Γ ) . (10.14)
vh ∈Sh
Proof. For σ = 0 and σ ∈ (0, 1] as well as for s ∈ [σ, 2] the approximation

property is just the error estimate (10.12) and (10.13), respectively.

As in Lemma 9.13 the L2 projection Qh : H 1/2 (Γ ) → Sh1 (Γ ) ⊂ H 1/2 (Γ )
defines a bounded operator satisfying
Qh vH 1/2 (Γ ) ≤ c vH 1/2 (Γ ) for all v ∈ H 1/2 (Γ ). (10.15)

Note that if the boundary decomposition is locally adaptive, to ensure (10.15)

we have to assume that the local mesh sizes of neighboring boundary elements
do no vary too strongly [137].
It remains to prove the inverse inequality of the trial space Sh0 (Γ ) of piece-
wise constant basis functions. For this we first define the global trial space
ShB (Γ ) of local bubble functions ϕB which are defined on τ . With respect to
the reference element τ ⊂ Rd−1 the associated form functions are given by

ξ(1 − ξ) for d = 2,
ψB (ξ) =
ξ1 ξ2 (1 − ξ1 − ξ2 ) for d = 3.
If the boundary decomposition is globally quasi–uniform there holds as in
(9.38) a global inverse inequality
vh H 1 (Γ ) ≤ cI h−1 vh L2 (Γ ) for all vh ∈ ShB (Γ ).

By using an interpolation argument we then conclude
vh H 1/2 (Γ ) ≤ cI h−1/2 vh L2 (Γ ) for all vh ∈ ShB (Γ ). (10.16)
h : L2 (Γ ) → Sh (Γ )
For a given u ∈ L2 (Γ ) we define the projection operator QB B
as the unique solution of the variational problem

(QBh u)(x)wh (x)ds x = u(x)wh (x)dsx for all wh ∈ Sh0 (Γ ). (10.17)
Γ Γ
When considering piecewise constant test functions this is equivalent to

(QBh u)(x)ds x = u(x)dsx for all = 1, . . . , N.
τ τ
As in Lemma 9.17 we can prove the stability estimate

√
QBh uL2 (Γ ) ≤ 2 uL2 (Γ ) for all u ∈ L2 (Γ ). (10.18)
Hence we can formulate an inverse inequality of the trial space Sh0 (Γ ) of
piecewise constant basis functions.
Lemma 10.10. Assume that the boundary decomposition (10.2) is globally
quasi–uniform. Then there holds the global inverse inequality
wh L2 (Γ ) ≤ cI h−1/2 wh H −1/2 (Γ ) for all wh ∈ Sh0 (Γ ).
Proof. For wh ∈ Sh0 (Γ ) we have by using (10.17)
wh , vL2 (Γ ) wh , QB
h vL2 (Γ )
wh L2 (Γ ) = sup = sup
0=v∈L2 (Γ ) vL2 (Γ ) 0=v∈L2 (Γ ) v L2 (Γ )
QB
h vH 1/2 (Γ )
≤ wh H −1/2 (Γ ) sup .
0=v∈L2 (Γ ) vL2 (Γ )
By applying the inverse inequality (10.16) as well as the stability estimate

(10.18) we finally obtain the assertion.

11
Finite Element Methods
For the approximate solution of the variational problems as described in Chap-

ter 4 we will use the finite–dimensional trial spaces which were constructed in
Chapter 9. Here we will just consider finite elements of lowest order, in partic-
ular we will use piecewise linear and continuous basis functions. The stability
and error analysis is imbedded in the general theory as given in Chapter 8.
Some numerical examples illustrate the theoretical results.

For the Poisson equation we consider the Dirichlet boundary value problem
(1.10) and (1.11),
−∆u(x) = f (x) for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ = ∂Ω. (11.1)
Let ug ∈ H 1 (Ω) be some bounded extension of the given Dirichlet datum

g ∈ H 1/2 (Γ ). Then the variational problem is to find u0 := u − ug ∈ H01 (Ω)
such that

∇u0 (x)∇v(x)dx = f (x)v(x)dx − ∇ug (x)∇v(x)dx (11.2)
Ω Ω Ω
is satisfied for all v ∈ H01 (Ω). By using Theorem 4.3 we can state the unique
solvability of the above variational formulation.
Let
-
Xh := Sh1 (Ω) ∩ H01 (Ω) = span{ϕ1i }M
i=1
be the conformal trial space of piecewise linear and globally continuous basis
functions ϕ1k which are zero on the boundary ∂Ω. Note that the trial space is
defined with respect to some admissible decomposition Ω = ∪N =1 τ of Ω into
finite elements τ . Then the Galerkin variational problem of (11.2) is to find
u0,h ∈ Xh such that
244 11 Finite Element Methods

∇u0,h (x)∇vh (x)dx = f (x)vh (x)dx − ∇ug (x)∇vh (x)dx (11.3)
Ω Ω Ω
is satisfied for all vh ∈ Xh . By applying Theorem 8.1 (Cea’s Lemma) there ex-
ists a unique solution u0,h of the Galerkin variational problem (11.3) satisfying
the error estimate (8.7),
cA
2
u0 − u0,h H 1 (Ω) ≤ inf u0 − vh H 1 (Ω) .
cA
1 vh ∈Xh
If the solution u of the Dirichlet boundary value problem (11.1) satisfies

u ∈ H s (Ω) for some s ∈ [1, 2], then we obtain by applying the trace the-
orem g = γ0int u ∈ H s−1/2 (Γ ), i.e. the extension ug of the given Dirichlet
datum g can be chosen such that ug ∈ H s (Ω) is satisfied. Therefore we have
u0 = u−ug ∈ H s (Ω) and from the approximation property (9.33) we conclude
the error estimate
u0 − u0,h H 1 (Ω) ≤ c hs−1 |u|H s (Ω) for u ∈ H s (Ω), s ∈ [1, 2]. (11.4)
When assuming certain smoothness properties of the domain Ω we can prove,

by using some duality arguments, error estimates which are valid in L2 (Ω).
Theorem 11.1 (Aubin–Nitsche Trick). Suppose that Ω is convex or the

boundary Γ = ∂Ω is smooth. For given f ∈ L2 (Ω) and g = γ0int ug with
ug ∈ H 2 (Ω) let u0 ∈ H01 (Ω) be the unique solution of the variational problem

∇u0 (x)∇v(x)dx = f (x)v(x)dx − ∇ug (x)∇v(x)dx
Ω Ω Ω
to be satisfied for all v ∈ H01 (Ω). Assume that

u0 H 2 (Ω) ≤ c f L2 (Ω) + ug H 2 (Ω) .
The approximate solution u0,h ∈ Xh of the Galerkin variational problem (11.3)

then satisfies the error estimate
0 1
u0 − u0,h L2 (Ω) ≤ c h2 f L2 (Ω) + ug H 2 (Ω) .
Proof. By assumption we have u0 ∈ H 2 (Ω), and for the approximate solution

u0,h ∈ Xh we get by using (11.4) the error estimate
0 1
u0 − u0,h H 1 (Ω) ≤ c h |u0 |H 2 (Ω) ≤ c h f L2 (Ω) + ug H 2 (Ω) . (11.5)
Let w ∈ H01 (Ω) be the unique solution of the variational problem

∇w(x)∇v(x)dx = [u0 (x) − u0,h (x)]v(x)dx
Ω Ω
to be satisfied for all v ∈ H01 (Ω). Due to the assumptions made on Ω we

conclude w ∈ H 2 (Ω), and
wH 2 (Ω) ≤ c u0 − u0,h L2 (Ω) .
Due to u0 − u0,h ∈ H01 (Ω) and by using the Galerkin orthogonality

∇[u0 (x) − u0,h (x)]∇vh (x)dx = 0 for all vh ∈ Xh
Ω
we get

u0 − u0,h 2L2 (Ω) = [u0 (x) − u0,h (x)][u0 (x) − u0,h (x)]dx
Ω

= ∇w(x)∇[u0 (x) − u0,h (x)]dx
Ω

= ∇[w(x) − Q1h w(x)]∇[u0 (x) − u0,h (x)]dx
Ω
≤ w − Q1h wH 1 (Ω) u0 − u0,h H 1 (Ω)
where Q1h : H01 (Ω) → Xh ⊂ H01 (Ω) is the H01 (Ω) projection which is defined
similar to (9.29). For w ∈ H 2 (Ω) we obtain from the approximation property
(9.32) of the trial space Xh the error estimate
w − Q1h wH 1 (Ω) ≤ c h wH 2 (Ω) ≤ c h u0 − u0,h L2 (Ω) .
Altogether we have
u0 − u0,h L2 (Ω) ≤ c h u0 − u0,h H 1 (Ω) ,
and by using (11.5) the assertion follows.

To realize the Galerkin variational formulation (11.3) we need to know the
bounded extension ug ∈ H 2 (Ω) of the given Dirichlet datum g. Formally, for
ug ∈ H 2 (Ω) we denote by Ih ug ∈ Sh1 (Ω) the piecewise linear interpolation
which can be written as

M
Ih ug (x) = ug (xi )ϕ1i (x) .
i=1
Instead of the exact Galerkin variational formulation (11.3) we now consider

a perturbed variational problem to find u 0,h ∈ Xh such that

∇
u0,h (x)∇vh (x)dx = f (x)vh (x)dx − ∇Ih ug (x)∇vh (x)dx (11.6)
Ω Ω Ω
is satisfied for all vh ∈ Xh . For the unique solution u 0,h ∈ Xh we have, by

using Theorem 8.2 (Strang Lemma), the error estimate

cA
2
u0 − u0,h H 1 (Ω) ≤ A inf u0 − vh H 1 (Ω) + ug − Ih ug H 1 (Ω) .
c1 vh ∈Xh
If we assume u ∈ H s (Ω) for some s ∈ [1, 2] and ug ∈ H 2 (Ω), then by using

the approximation property of the trial space Xh as well as the interpolation
estimate (9.22) we get
u0 − u
0,h H 1 (Ω) ≤ c1 hs−1 |u0 |H s (Ω) + c2 h |ug |H 2 (Ω)
0 1
≤ c hs−1 |u0 |H s (Ω) + gH 3/2 (Γ ) .
The piecewise linear interpolation Ih ug can be written, by considering the

-
interior nodes {xi }Mi=1 , xi ∈ Ω and the boundary nodes {xi }i=M
M
-+1 , xi ∈ Γ ,
as
-
M
M
Ih ug (x) = ug (xi )ϕ1i (x) + g(xi )ϕ1i (x).
i=1 -+1
i=M
Hence the perturbed Galerkin variaitional formulation (11.6) is equivalent to

finding
- -

M
M
0,h (x) +
ū0,h (x) := u Ih ug (xi )ϕ1i (x) =: ūi ϕ1i (x) ∈ Xh
i=1 i=1
as the unique solution of the variational problem

M
∇ū0,h (x)∇vh (x)dx = f (x)vh (x)dx − g(xi ) ∇ϕ1i (x)∇vh (x)dx
Ω Ω -+1
i=M Ω
(11.7)
to be satisfied for all vh ∈ Xh . The resulting approximate solution of the
Dirichlet boundary value problem is then given by
-

M
M
uh (x) := ūk ϕ1i (x) + g(xi )ϕ1i (x) ∈ Sh1 (Ω). (11.8)
i=1 -+1
i=M
Theorem 11.2. Let u ∈ H s (Ω) for some s ∈ [1, 2] be the unique solution of
the Dirichlet boundary value problem (11.1). Let ug ∈ H 2 (Ω) be an appropri-
ate chosen extension of the given Dirichlet datum g. Let uh be the approximate
solution of (11.1) which is defined by (11.8). Then there holds the error esti-
mate 0 1
u − uh H 1 (Ω) ≤ c hs−1 |u0 |H s (Ω) + ug H 2 (Ω) (11.9)
Proof. By using the triangle inequality we have

u − uh H 1 (Ω) = u0 + ug − (
u0,h + Ih ug )H 1 (Ω)
≤ u0 − u
0,h H 1 (Ω) + ug − Ih ug H 1 (Ω) .
Therefore, the assertion follows from Theorem 8.2, by estimating the interpo-
lation error, and by applying the inverse trace theorem.
If the solution u ∈ H 2 (Ω) is sufficiently regular, we therefore obtain the
error estimate
u − uh H 1 (Ω) ≤ c h |u|H 2 (Ω) . (11.10)
As in Theorem 11.1 (Aubin–Nitsche Trick) we can also prove an estimate for
the error u0 − u
0,h L2 (Ω) .
Lemma 11.3. Let all assumptions of Theorem 11.1 be valid, in particular we
assume u ∈ H 2 (Ω) to be the unique solution of the Dirichlet boundary value
problem (11.1). Then there holds the error estimate
0 1
0,h L2 (Ω) ≤ c h2 |u0 |H 2 (Ω) + |ug |H 2 (Ω) .
u0 − u (11.11)
Proof. Let w ∈ H01 (Ω) be the unique solution of the variational problem

∇w(x)∇v(x)dx = [u0 (x) − u 0,h (x)]v(x)dx
Ω Ω
to be satisfied for all v ∈ H01 (Ω). For u0 − u

0,h ∈ L2 (Ω) and due to the
assumptions made on Ω we conclude w ∈ H 2 (Ω), and therefore we have
wH 2 (Ω) ≤ c u0 − u
0,h L2 (Ω) .
By subtracting the perturbed variational problem (11.6) from the variational
problem (11.2) this gives

∇[u0 (x) − u0,h (x)]∇vh (x)dx = ∇[Ih ug (x) − ug (x)]∇vh (x)dx
Ω Ω
for all vh ∈ Xh . Hence we have

2
u0 − u
0,h L2 (Ω) = [u0 (x) − u
0,h (x)][u0 (x) − u
0,h (x)]dx
Ω

= ∇w(x)∇[u0 (x) − u
0,h (x)]dx
Ω

= ∇[w(x) − Q1h w(x)]∇[u0 (x) − u
0,h (x)]dx
Ω

+ ∇Q1h w(x)∇[Ih ug (x) − ug (x)]dx.
Ω
The first term can be estimated as in the proof of Theorem 11.1,

∇[w(x) − Q1h w(x)]∇[u0 (x) − u0,h (x)]dx
Ω
≤ c h u0 − u
0,h L2 (Ω) u0 − u
0,h H 1 (Ω) .
For the second term we first have

∇Q1h w(x)∇[Ih ug (x) − ug (x)]dx
Ω

= ∇[Q1h w(x) − w(x)]∇[Ih ug (x) − ug (x)]dx
Ω

+ ∇w(x)∇[Ih ug (x) − ug (x)]dx.
Ω
The resulting first term can further be estimated by

∇[Q1h w(x) − w(x)]∇[Ih ug (x) − ug (x)]dx

Ω
1
≤ cA
2 w − Qh wH 1 (Ω) ug − Ih ug H 1 (Ω)
≤ c h wH 2 (Ω) ug − Ih ug H 1 (Ω)
≤ c h u0 − u
0,h L2 (Ω) ug − Ih ug H 1 (Ω) .
For the remaining second term we have, by applying integration by parts, for
w ∈ H01 (Ω) ∩ H 2 (Ω),

∇w(x)∇[Ih ug (x) − ug (x)]dx = − ∆w(x)[Ih ug (x) − ug (x)]dx

Ω Ω
≤ wH 2 (Ω) ug − Ih ug L2 (Ω)
≤ c u0 − u
0,h L2 (Ω) ug − Ih ug L2 (Ω) .
Altogether we have
u0 − u
0,h L2 (Ω)
0 1
0,h H 1 (Ω) + ug − Ih ug H 1 (Ω) + ug − Ih ug L2 (Ω) .
≤ c1 h u0 − u
For u0 , ug ∈ H 2 (Ω) the assertion now follows from the error estimates
for u0 − u0,h H 1 (Ω) as well as from the interpolation error estimates for
ug − Ih ug L2 (Ω) and ug − Ih ug H 1 (Ω) .

The Galerkin variational problem (11.7) to find the coefficients ūi for
- is equivalent to an algebraic system of linear equations Ah ū = f
i = 1, . . . , M
with the stiffness matrix Ah defined by

Ah [j, i] = ∇ϕ1i (x)∇ϕ1j (x)dx
Ω
-, and with the right hand side vector f given by

for i, j = 1, . . . , M

M
fj = f (x)ϕ1j (x)dx − g(xi ) ∇ϕ1i (x)∇ϕ1j (x)dx
Ω -+1
i=M Ω
-. The stiffness matrix Ah is symmetric and due to

for j = 1, . . . , M

2
(Ah v, v) = ∇vh (x)∇vh (x)dx ≥ cA
1 vh H 1 (Ω)
Ω
-
for all v ∈ RM ↔ vh ∈ Xh ⊂ H01 (Ω) positive definite. In particular we have
the following result:
-
Lemma 11.4. For all v ∈ RM ↔ vh ∈ Xh ⊂ H01 (Ω) there hold the spectral
equivalence inequalities
c1 hdmin v22 ≤ (Ah v, v) ≤ c2 hd−2 2

max v2 (11.12)
-
Proof. For v ∈ RM ↔ vh ∈ Xh ⊂ H01 (Ω) it follows by localization and by
applying the local inverse inequality (9.15)
- M
- - M
-

M
M
(Ah v, v) = Ah [j, i]vi vj = a(ϕ1i , ϕ1j )vi vj
i=1 j=1 i=1 j=1
- -

M
M
= a( vi ϕ1i , vj ϕ1j ) = a(vh , vh )
i=1 j=1
N

2
= |∇vh (x)| dx = |∇vh (x)|2 dx
Ω =1 τ

N
N
= ∇vh 2L2 (τ ) ≤ cI h−2 2
vh L2 (τ )
=1 =1
⎛ ⎞
-

N
M
≤c h−2 vk2 = c ⎝ hd−2 ⎠ vk2
∆
=1 k∈J( ) k=1 ∈I(k)
and therefore the upper estimate. On the other hand, by using the H01 (Ω)
ellipticity of the bilinear form a(·, ·) and by changing to the L2 (Ω) norm we
get
2 2
(Ah v, v) = a(vh , vh ) ≥ cA
1 vh H 1 (Ω) ≥ c1 vh L2 (Ω)
A
⎛ ⎞
-

N
N
M
= cA
1 vh 2L2 (τ ) ≥ c ∆ vk2 = c ⎝ hd ⎠ vk2
=1 =1 k∈J( ) k=1 ∈I(k)
and therefore the lower estimate.

Note that the constants in the spectral equivalence inequalities (11.12) are
sharp, i.e. the constants can not be improved. Hence we have for the spectral
condition number of the stiffness matrix Ah in the case of a globally quasi–
uniform mesh
κ2 (Ah ) ≤ c h−2 , (11.13)
in particular the spectral condition number increases when the mesh is re-
fined. As an example we consider a Dirichlet boundary value problem where
the domain is given by the square Ω = (0, 0.5)2 . The initial mesh consists of
four finite elements with five nodes, which are recursively refined by decom-
posing each finite element into four congruent elements, see Fig. 11.1 for the
refinement levels L = 0 and L = 3.
@ @ @ @ @ @ @ @ @
@ @@ @@ @@ @@
@ @ @@ @@ @@ @
@
@ @ @ @
@ @ @ @@ @@ @
@ @ @ @ @ @
@ @ @ @ @@ @@ @
@
@ @ @
@ @ @ @ @ @@ @
@ @ @ @ @ @ @
@ @ @ @ @ @ @@ @
@
@ @
@ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @
@ @
@ @ @ @ @ @ @ @
Fig. 11.1. Initial mesh (L = 0) and refined mesh L = 3.
The minimal and maximal eigenvalues and the resulting spectral condition
numbers of the associated finite element stiffness matrices are given in Table
11.1. Note that when choosing d = 2 and h = O(N −2 ) the results of Lemma
11.4 are confirmed. Thus, when using a conjugate gradient scheme to solve
the linear equation systems Ah u = f with the symmetric and positive def-
inite system matrices Ah we need to have an appropriate preconditioner to
bound the number of necessary iteration steps to reach a prescribed accuracy
independent of the system size. Note that preconditioned iterative schemes
are considered later in Chapter 13.
L N λmin (Ah ) λmax (Ah ) κ2 (Ah )

2 64 5.86 –1 7.41 12.66
3 256 1.52 –1 7.85 51.55
4 1024 3.84 –2 7.96 207.17
5 4096 9.63 –3 7.99 829.69
6 16384 2.41 –3 8.00 3319.76
7 65536 6.02 –4 8.00 13280.04
8 262144 1.52 –4 8.00 52592.92
Theory: O(h2 ) O(1) O(h−2 )
Table 11.1. Spectral condition numbers of the stiffness matrices Ah .
Next we will discuss the computation of the load vector f and the realization
of a matrix by vector multiplication with the stiffness matrix Ah as needed in
- we have by
the application of an iterative solution scheme. For j = 1, . . . , M
localization and parametrization, see Chapter 9,

fj = f (x)ϕ1j (x)dx = f (x)ϕ1j (x)dx
Ω ∈I(j) τ

= |det J | f (x 1 + J ξ)ψ 1j (ξ)dξ
∈τj τ
where j is the local index of the global node xj with respect to the finite
element τ . Hence we can reduce the computation of the global load vector
f to the computation of local load vectors f . In particular, for each finite
element τ we need to compute

f ,ι = |det J | f (x 1 + J ξ)ψι1 (ξ)dξ for ι = 1, . . . , d + 1.
τ
For ι = 1, . . . , d + 1 we denote by ι the corresponding global node index, then

the global load vector f is computed by assembling all local load vectors f ,
i.e.
f ι := f ι + f ,ι .
-
If u ∈ RM ↔ uh ∈ Xh is given, the result v = Ah u of a matrix by vector
multiplication with the global stiffness matrix Ah can be written as
-

M
vj = Ah [j, i]ui = a(uh , ϕ1j ) = ∇uh (x)∇ϕ1j (x)dx
i=1 ∈I(j) τ

= ui ∇ϕ1i (x)∇ϕ1j (x)dx, -.
j = 1, . . . , M
∈I(j) i∈J( ) τ
Therefore it is sufficient to compute the local stiffness matrices A h which are

defined by

A h [ι , ι] = |det J | J − ∇ξ ψι1 (ξ)J − ∇ξ ψι1 (ξ)dξ
τ
for ι, ι = 1, . . . , d+1. Hence we can reduce the matrix by vector multiplication

with the global stiffness matrix Ah to a localization of the global degrees of
-
freedom u ∈ RM to local degrees of freedom u ∈ Rd+1 where u ,ι = u ι , a
multiplication with the local stiffness matrices v = A h u , and the assembling
of the global vector v from the local results v , i.e.
v ι := v ι + v ,ι .
The incorporation of given Dirichlet boundary conditions to compute the

global load vector f can be done in the same way. Hence, for a matrix by
vector multiplication with the global stiffness matrix Ah we only need to
store the local stiffness matrices A h with an effort of O(N ) essential opera-
tions. For alternative approaches to describe the sparse stiffness matrix Ah ,
see, for example, [85].
To check the theoretic error estimates (11.9) and (11.11) we now consider
the Dirichlet boundary value problem (11.1) where Ω = (0, 0.5)2 and f = 0
are given, and where the Dirichlet boundary data are prescribed such that
1
u(x) = − log |x − x∗ |, x∗ = (−0.1, −0.1) (11.14)
2
is the exact solution. In Table 11.2 we give the errors of the approximate
Galerkin solutions uh for a sequence of uniformly refined meshes, where L is
the refinement level, N is the number of finite elements, M is the total number
of nodes, and DoF is the number of degrees of freedom which coincides with
the number of interior nodes.
L N M DoF |u − uh |H 1 (Ω) eoc u − uh L2 (Ω) eoc

2 64 41 25 1.370 –1 2.460 –3
3 256 145 113 6.954 –2 0.98 5.717 –4 2.11
4 1024 545 481 3.494 –2 0.99 1.408 –4 2.02
5 4096 2113 1985 1.749 –2 1.00 3.511 –5 2.00
6 16384 8321 8065 8.748 –3 1.00 8.771 –6 2.00
7 65536 33025 32513 4.374 –3 1.00 2.192 –6 2.00
8 262144 131585 130561 2.187 –3 1.00 5.481 –7 2.00
9 1048576 525313 523265 1.094 –3 1.00 1.370 –7 2.00
Theory: 1 2
Table 11.2. Errors and estimated order of convergence, Dirichlet problem.

Since the solution (11.14) of the Dirichlet boundary value problem is infinitely
often differentiable, we can apply Theorem 11.2 and Lemma 11.3 for s = 2.
Hence we obtain one as order of convergence when measuring the error in the
energy norm |u − uh |H 1 (Ω) . Moreover, when applying the Aubin–Nitsche trick
we get two as order of convergence when measuring the error in the L2 norm
u − uh L2 (Ω) . By eoc we denote the estimated order of convergence which
can be computed from
log u − uh − log u − uh+1

eoc := .
log h − log h +1
Note that the theoretical error estimates are well confirmed by the numerical
results as documented in Table 11.2.

We now consider the Neumann boundary value problem (1.10) and (1.12) for
the Poisson equation,
−∆u(x) = f (x) for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ = ∂Ω (11.15)
where we have to assume the solvability condition (1.17)

f (x) dx + g(x) dsx = 0. (11.16)
Ω Γ
For a finite element discretization we consider the modified variational prob-

lem (4.31) which admits, due to Section 4.1.3, a unique solution u ∈ H 1 (Ω)
such that

∇u(x)∇v(x)dx+ u(x)dx v(x)dx = f (x)v(x)dx+ g(x)γ0int v(x)dsx
Ω Ω Ω Ω Γ
(11.17)
is satisfied for all v ∈ H 1 (Ω). From the solvability condition (11.16) we
then conclude the scaling condition u ∈ H∗1 (Ω), i.e. u ∈ H 1 (Ω) satisfying
u, 1L2 (Ω) = 0.
Let
Xh := Sh1 (Ω) = span{ϕ1k }M 1
k=1 ⊂ H (Ω)
be the conforming trial space of piecewise linear and globally continuous basis
functions ϕ1k with respect to an admissible finite element mesh Ω = ∪N =1 τ .
Note that the basis functions {ϕ1k }M
k=1 build a partition of unity, i.e.

M
ϕ1k (x) = 1 for all x ∈ Ω. (11.18)
k=1
The Galerkin variational formulation of (11.17) is to find uh ∈ Xh such that

∇uh (x)∇vh (x)dx + uh (x)dx vh (x)dx (11.19)
Ω Ω Ω

= f (x)vh (x)dx + g(x)vh (x)dsx
Ω Γ
is satisfied for all vh ∈ Xh . By applying Theorem 8.1 (Cea’s Lemma) we

conclude that there exists a unique solution uh of the variational problem
(11.19) satisfying the error estimate (8.7),
ĉA
1
u − uh H 1 (Ω) ≤ inf u − vh H 1 (Ω) .
cA2 vh ∈Xh
If the solution u of the Neumann boundary value problem (11.15) satisfies

u ∈ H s (Ω) for some s ∈ [1, 2], then we conclude, by using the approximation
property (9.33) the error estimate
u − uh H 1 (Ω) ≤ c hs−1 |u|H s (Ω) for u ∈ H s (Ω), s ∈ [1, 2].
Due to (11.18) we can choose vh ≡ 1 as a test function of the Galerkin

variational formulation (11.19). From the solvability condition (11.16) we then
obtain
uh (x)dx dx = 0
Ω Ω
and therefore uh ∈ H∗1 (Ω),i.e. the scaling condition is automatically satisfied

for the Galerkin solution uh ∈ Xh .
The Galerkin variational problem (11.19) to find the coefficient vector
u ∈ RM is equivalent to the solution of the linear system of algebraic equations
0 1
Ah + a a u = f
where the stiffness matrix Ah defined by

Ah [j, i] = ∇ϕ1i (x)∇ϕ1j (x)dx
Ω
for i, j = 1, . . . , M , and with the load vector given by

fj := f (x)ϕ1j (x)dx + g(x)ϕ1j (x)dsx
Ω Γ
for j = 1, . . . , M . In addition, a ∈ RM is defined by

ai = ϕ1i (x) dx
Ω
11.3 Finite Element Methods with Lagrange Multipliers 255
for i = 1, . . . , M . Note that the modified stiffness matrix Ah + a a is sym-

metric and positive definite. Moreover, the spectral equivalence inequalities
(11.12) remain valid. Hence, when using the conjugate gradient scheme to
solve the linear system iteratively we have to use again an appropriate precon-
ditioner. Both the computation of the load vector f as well as an application
of the matrix by vector product with the stiffness matrix Ah can be realized
as for the Dirichlet boundary value problem.
The approximate solution of boundary value problems with mixed bound-
ary conditions as well as with Robin boundary conditions can be formulated
and analyzed as for the Dirichlet or Neumann boundary value problem. Here,
we will not discuss this in detail. The same is true for the approximate solution
of boundary value problems in linear elasticity. However, when considering the
Neumann boundary value problem we have to modify both the solvability con-
ditions as well as the definition of the modified variational problem due to the
rigid body motions.
11.3 Finite Element Methods with Lagrange Multipliers

For an alternative approximation of the Dirichlet boundary value problem
(11.1) we consider the modified saddle point problem (4.22) and (4.23) to find
(u, λ) ∈ H 1 (Ω) × H −1/2 (Γ ) such that

γ0int u(x)dsx γ0int v(x)dsx + ∇u(x)∇v(x)dx − γ0int v(x)λ(x)dsx
Γ Γ Ω Γ

= f, vΩ + g(x)dsx γ int v(x)ds
0 x (11.20)
Γ Γ

γ0int u(x)µ(x)dsx + λ(x)dsx µ(x)dsx
Γ Γ
Γ
= g, µΓ − f (x)dx µ(x)dsx
Ω Γ
is satisfied for all (v, µ) ∈ H 1 (Ω) × H −1/2 (Γ ).

Assume that there is given an admissible finite element mesh Ω = ∪N =1 τ
Ω
of the polygonal or polyhedral bounded domain Ω ⊂ R . The restriction of

d
the finite element mesh in Ω defines a boundary element mesh Γ = ∪N =1 Γ

Γ
on Γ = ∂Ω. Let
Xh (Ω) := Sh1 (Ω) = span{ϕ1i }M 1

i=1 ⊂ H (Ω)
be the conforming finite element space of piecewise linear and globally contin-
uous basis functions ϕ1i . The restriction of Xh (Ω) onto Γ = ∂Ω then defines
a boundary element space
Xh (Γ ) := Sh1 (Γ ) = span{φ1k }M
i=1 ⊂ H
Γ 1/2
(Γ )
of piecewise linear and continuous basis functions φ1k . We denote by MΩ the
number of all interior nodes xi ∈ Ω, and we have M = MΩ + MΓ as well as
Xh (Ω) = span{ϕ1i }M 1 M
i=1 ∪ span{ϕi }i=MΩ +1 .
Ω
In particular,
φ1i = γ0int ϕ1MΩ +i for i = 1, . . . , MΓ .
Moreover, let
−1/2
ΠH := span{ψk }N
k=1 ⊂ H (Γ )
denote a suitable trial space to approximate the Lagrange multiplier λ.
The Galerkin discretization of the saddle point problem (11.20) is to find
(uh , λH ) ∈ Xh × ΠH such that

γ0 uh (x)dsx γ0 vh (x)dsx + ∇uh (x)∇vh (x)dx − γ0int vh (x)λH (x)dsx
int int
Γ Γ Ω Γ

= f, vh Ω + g(x)dsx γ0int vh (x)dsx (11.21)
Γ Γ

γ0int uh (x)µH (x)dsx + λH (x)dsx µH (x)dsx
Γ Γ Γ

= g, µH Γ − f (x)dx µH (x)dsx
Ω Γ
is satisfied for all (vh , µH ) ∈ Xh × ΠH . With

Ah [j, i] := ∇ϕ1i (x)∇ϕ1j (x)dx,
Ω

Bh [, i] := ψ (x)γ0int ϕ1i (x)dsx ,
Γ

ai := γ0int ϕ1i (x)dsx ,
Γ

b := ψ (x)dsx
Γ
for i, j = 1, . . . , M and = 1, . . . , N , as well as with

fj := f (x)ϕ1j (x)dx + g(x)dsx γ0int ϕ1j (x)dsx ,
Ω Γ Γ

g := g(x)ψ (x)dsx − f (x)dx ψ (x)dsx
Γ Ω Γ
for j = 1, . . . , M and = 1, . . . , N we conclude that the approximate saddle

point problem (11.21) is equivalent to a linear system of algebraic equations,
& '& ' & '
a a + Ah −Bh u f
= . (11.22)
Bh bb λ g
Obviously,

0 for i = 1, . . . , MΩ ,
Bh [, i] =
B̄h [, i − MΩ ] for i = MΩ + 1, . . . , M
with
B̄h [, i] = ψ (x)φ1i (x)dsx
Γ
for i = 1, . . . , MΓ and = 1, . . . , N .
The matrix a a + Ah is by construction symmetric and positive definite
and therefore invertible. In particular, the first equation in (11.22) can be
solved for u to obtain
0 1−1 0 1
u = a a + Ah f + Bh λ .
Inserting this into the second equation of (11.22) we end up with the Schur
complement system
2 0 1−1 3 0 1−1
Bh a a + Ah Bh + b b λ = g − Bh a a + Ah f. (11.23)
The unique solvability of the Schur complement system (11.23) and therefore
of the linear system (11.22), and hence of the discrete saddle point problem
(11.21), now follows from Lemma 8.6 where we have to ensure the discrete
stability condition (8.25), i.e.
µH , γ0int vh Γ
cS µH H −1/2 (Γ ) ≤
sup for all µH ∈ ΠH . (11.24)
0=vh ∈Xh (Ω) vh H 1 (Ω)
This stability condition is first considered for the boundary element trial
spaces ΠH and Xh (Γ ).
Theorem 11.5. The mesh size h of the trial space Xh (Γ ) is assumed to be

sufficiently small compared to the mesh size H of ΠH , i.e. h ≤ c0 H. For the
trial space ΠH we assume a global inverse inequality. Then there holds the
stability condition
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ sup for all µH ∈ ΠH . (11.25)
0=wh ∈Xh (Γ ) wh H 1/2 (Γ )
Proof. Let µH ∈ ΠH ⊂ H −1/2 (Γ ) be arbitrary but fixed. By using the Riesz

representation theorem (Theorem 3.3) there exists a unique JµH ∈ H 1/2 (Γ )
satisfying
JµH , wH 1/2 (Γ ) = µH , wΓ for all w ∈ H 1/2 (Γ )
1/2
and JµH H 1/2 (Γ ) = µH H −1/2 (Γ ) . Let Qh JµH ∈ Xh (Γ ) be the unique
1/2
Qh JµH , wh H 1/2 (Γ ) = µH , wh Γ for all wh ∈ Xh (Γ ).
Then there holds the error estimate

1/2
(I − Qh )JµH H 1/2 (Γ ) ≤ inf JµH − wh H 1/2 (Γ ) .
wh ∈Xh (Γ )
Due to µH ∈ L2 (Γ ) we obtain, by using duality and the definition of JµH ,
JµH , vH 1/2 (Γ ) µH , vΓ

JµH H 1 (Γ ) = sup = sup ≤ µH L2 (Γ )
0=v∈L2 (Γ ) v L 2 (Γ ) 0=v∈L2 (Γ ) vL2 (Γ )
and therefore JµH ∈ H 1 (Γ ). From the approximation property of the trial

space Xh (Γ ), see the error estimate (10.13), we then obtain
1/2
(I − Qh )JµH H 1/2 (Γ ) ≤ cA h1/2 JµH H 1 (Γ ) ≤ cA h1/2 µH L2 (Γ ) .
By applying the inverse inequality in the trial space ΠH this gives

1/2
1/2 h
(I − Qh )JµH H 1/2 (Γ ) ≤ cA cI µH H −1/2 (Γ ) .
H
Assume that the constant c0 is chosen such that h ≤ c0 H and
1/2 1
(I − Qh )JµH H 1/2 (Γ ) ≤ µH H −1/2 (Γ )
2
is satisfied. Then we have
µH H −1/2 (Γ ) = JµH H 1/2 (Γ )

1/2 1/2
≤ Qh JµH H 1/2 (Γ ) + JµH − Qh JµH H 1/2 (Γ )
1/2 1
≤ Qh JµH H 1/2 (Γ ) + µH H −1/2 (Γ )
2
and therefore
1/2 1
Qh JµH H 1/2 (Γ ) ≥ µH H −1/2 (Γ ) .
2
1/2
By using the definition of Qh JµH we get
1/2 1/2 1/2
µH , Qh JµH Γ = Qh JµH , Qh JµH H 1/2 (Γ )
1/2
= Qh JµH 2H 1/2 (Γ )
1 1/2
≥ Qh JµH H 1/2 (Γ ) µH H −1/2 (Γ )
2
from which the stability estimate (11.25) follows.

Remark 11.6. To establish the stability condition (11.25) one may also use
Fortin’s criterion (Lemma 8.9). Then we have to prove the boundedness of
the projection operator Q h : H 1/2 (Γ ) → Xh (Γ ) ⊂ H 1/2 (Γ ) which is defined
via the variational formulation
h u, µH Γ = u, µH Γ
Q for all µH ∈ ΠH .
If we assume a global inverse inequality in the trial space Xh (Γ ), then we can

prove the H 1/2 (Γ )–boundedness of Q h by using the error estimates of I − Qh
1/2 1/2 1/2
and I − Qh in L2 (Γ ) as well as the stability of Qh in H (Γ ). For the
case of trial spaces which are defined with respect to some adaptive boundary
element mesh where we can not assume an inverse inequality globally, we refer
to [137, 138].
By using the inverse trace theorem (Theorem 2.22) the stability condition
(11.25) implies
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ cIT sup for all µH ∈ ΠH .
0=wh ∈Xh (Γ ) Ewh H 1 (Ω)
Finally, let Rh : H 1 (Ω) → Xh (Ω) ⊂ H 1 (Ω) be some quasi interpolation

operator [133] satisfying
Rh vH 1 (Ω) ≤ cR vH 1 (Ω)
and where Dirichlet boundary conditions are preserved. Then we obtain
µH , wh Γ
c̄S µH H −1/2 (Γ ) ≤ cIT cR sup for all µH ∈ ΠH ,
0=wh ∈Xh (Γ ) Rh Ewh H 1 (Ω)
and by choosing vh = Rh Ewh ∈ Xh (Ω) we conclude the stability condition

(11.24). This gives us the unique solvability of the Schur complement system
(11.23) and therefore of the linear system (11.22). The application of Theorem
8.8 yields, when assuming u ∈ H 2 (Ω) and λ ∈ Hpw 1
(Γ ), the error estimate
u−uh 2H 1 (Ω) +λ−λH 2H −1/2 (Γ ) ≤ c1 h2 |u|2H 2 (Ω) +c2 H 3 λ2Hpw

1 (Γ ) . (11.26)
To ensure the discrete stability condition (11.25) we need to assume h ≤ c0 H

where the constant c0 < 1 is sufficiently small.
We now consider the numerical example of Section 11.1 with h = 12 H. For

L = 3 the finite element mesh of Ω and the associated boundary element mesh
are depicted in Fig. 11.2. In Table 11.3 the computed errors of the approximate
solutions (uh , λH ) ∈ Sh1 (Ω) × SH
0
(Γ ) are given. The numerical results for the
approximate solution uh of the primal variable u confirm the theoretical error
estimate (11.26).
t t t t t
@ @ @ @ @ @ @ @ @ @
@ @
@ @ @ @
@
@ @ @@ @ @ @@
t @ @ @ @t
@ @ @@ @ @
@ @ @ @ @ @
@ @ @
@ @ @ @ @ @ @@
t@ @ @ @ @ @ @t
@ @ @ @ @ @ @@
@ @ @ @ @ @ @
@ @ @ @ @ @ @@
@
t @ @ @ @ @ @ @t
@ @
@ @ @ @ @ @ @ @
@ @ @ @ @ @ @ @
t @ @t @ @t @ @t @ @t
Fig. 11.2. Finite and boundary element meshes of Ω (L = 3).
L NΩ NΓ u − uh H 1 (Ω) eoc λ − λh L2 (Γ ) eoc

1 16 4 5.051 –1 1.198 ±0
2 64 8 2.248 –1 1.17 9.350 –1 0.36
3 256 16 9.897 –2 1.18 5.662 –1 0.72
4 1024 32 4.133 –2 1.26 2.970 –1 0.93
5 4096 64 1.851 –2 1.16 1.457 –1 1.03
6 16384 128 8.889 –3 1.06 7.093 –2 1.04
7 65536 256 4.393 –3 1.02 3.482 –2 1.03
8 262144 512 2.190 –3 1.00 1.723 –2 1.01
9 1048576 1024 1.094 –3 1.00 8.569 –3 1.01
Theory: 1 0.5
Table 11.3. Results for a Finite Element Method with Lagrange multipliers.
To describe the error of the approximation λH of the Lagrange multiplier λ

we use the L2 norm which is easier to compute.
0
Lemma 11.7. For the trial space SH (Γ ) we assume a global inverse inequality
2 1
to be valid. If u ∈ H (Ω) and λ ∈ Hpw (Γ ) are satisfied, then there holds the
error estimate
11.4 Exercises 261
λ − λH 2L2 (Γ ) ≤ c1 h2 H −1 |u|2H 2 (Ω) + c2 H 2 λ2Hpw

1 (Γ ) .
1 0
Proof. For λ ∈ Hpw (Γ ) we define QH λ ∈ SH (Γ ) to be the L2 projection as
defined in (10.4). By using the triangle inequality and the inverse inequality
we obtain
λ − λH 2L2 (Γ ) ≤ 2 λ − QH λ2L2 (Γ ) + 2 QH λ − λH 2L2 (Γ )

≤ 2 λ − QH λ2L2 (Γ ) + 2 c2I H −1 QH λ − λH 2H −1/2 (Γ )
2 3
≤ 2 λ − QH λ2L2 (Γ ) + 4 c2I H −1 QH λ − λ2H −1/2 (Γ ) + λ − λH 2H −1/2 (Γ ) .
The error estimate now follows from Theorem 10.2, Corollary 10.3, and by
using the error estimate (11.26).

When choosing h = 12 H we conclude by applying Lemma 11.7 the error
estimate
1
λ − λH 2L2 (Γ ) ≤ c1 H |u|2H 2 (Ω) + c2 H 2 λ2Hpw
1 (Γ )
4
and therefore an asymptotic order of convergence which is 0.5 when measur-
ing the error in the L2 norm. However, the numerical results in Table 11.3
indicate a higher order of convergence which is equal to 1. This preasymptotic
behavior may be explained by different orders of magnitude in the constants
1 2 2
4 c1 uH 1 (Ω) and c2 tH 1 (Γ ) .
pw
11.4 Exercises
11.1 Consider the Dirichlet boundary value problem
−u (x) = f (x) for x ∈ (0, 1), u(0) = u(1) = 0.
Compute the finite element stiffness matrix when using piecewise linear basis
functions with respect to a uniform decomposition of the interval (0, 1).
11.2 Show that the eigenvectors of the finite element stiffness matrix as de-
rived in Exercise 11.1 are given by the nodal interpolation of the eigenfunctions
as obtained in Exercise 1.6. Compute the associated eigenvalues and discuss
the behavior of the resulting spectral condition number.
11.3 Derive a two–dimensional Gaussian quadrature formula which integrates
cubic polynomials over the reference triangle
τ = {x ∈ R2 : x1 ∈ (0, 1), x2 ∈ (0, 1 − x1 )}
exactly.
12
Boundary Element Methods
The solution of the scalar homogeneous partial differential equation
Lu(x) = 0 for x ∈ Ω
is given by the representation formula
u(x) = (V γ1int u)(x) − (W γ0int u)(x) for x ∈ Ω.
Since the Cauchy data [γ0int u(x), γ1int u(x)] are given by the boundary con-
ditions only partially for x ∈ Γ , we need to find the remaining data from
the solution of suitable boundary integral equations. This chapter describes
boundary element methods as numerical discretization schemes to solve these
boundary integral equations approximately.

We first consider the Dirichlet boundary value problem (7.6),
Lu(x) = 0 for x ∈ Ω, γ0int u(x) = g(x) for x ∈ Γ. (12.1)
The solution u of (12.1) is given by the representation formula (6.1),

u(
x) = U ∗ ( int U ∗ (
x, y)γ1int u(y)dsy − γ1,y ∈ Ω. (12.2)
x, y)g(y)dsy , x
Γ Γ
We have to find the yet unknown conormal derivative t := γ1int u ∈ H −1/2 (Γ )

as the unique solution of the variational formulation (7.9),
1
V t, τ Γ = ( I + K)g, τ Γ for all τ ∈ H −1/2 (Γ ). (12.3)
2
Let
264 12 Boundary Element Methods
Sh0 (Γ ) = span{ϕ0k }N
k=1 ⊂ H
−1/2
(Γ )
be the trial space of piecewise constant basis functions ϕ0k . By using

N
th (x) = tk ϕ0k (x) ∈ Sh0 (Γ ) (12.4)
k=1
the Galerkin variational formulation of (12.3) reads to find th ∈ Sh0 (Γ ) such

that
1
V th , τh Γ = ( I + K)g, τh Γ (12.5)
2
is satisfied for all τh ∈ Sh0 (Γ ).
Note that the single layer potential V : H −1/2 (Γ ) → H 1/2 (Γ ) is bounded
(see (6.8)) and H −1/2 (Γ )–elliptic (see Theorem 6.22 in the three–dimensional
case d = 3, and Theorem 6.23 for d = 2 where we assume diam Ω < 1).
Moreover, Sh0 (Γ ) ⊂ H −1/2 (Γ ) is a conforming trial space. Therefore, all as-
sumptions of Theorem 8.1 (Cea’s Lemma) are satisfied, in particular there
exists a unique solution th ∈ Sh0 (Γ ) of the Galerkin variational formulation
(12.5) satisfying the stability estimate
1 1 cW
2
th H −1/2 (Γ ) ≤ ( I + K)gH 1/2 (Γ ) ≤ gH 1/2 (Γ ) ,
cV1 2 cV1

cV2
t − th H −1/2 (Γ ) ≤ inf t − τh H −1/2 (Γ ) . (12.6)
cV1 0 (Γ )
τh ∈Sh
In the case of a piecewise smooth Lipschitz boundary Γ = ∂Ω with the repre-

sentation Γ = ∪Jj=1 Γ j we obtain, by applying Lemma 2.20 and by using the
local definition of the trial space Sh0 (Γ ), the error estimate
cV2
J
t − th H −1/2 (Γ ) ≤ inf t|Γj − τhj H
-−1/2 (Γj ) . (12.7)
cV1 j=1 τhj ∈Sh0 (Γj )
From the approximation property (10.10) we then obtain the a priori error
estimate
1
t − th H −1/2 (Γ ) ≤ c hs+ 2 |t|Hpw
s (Γ ) (12.8)
when assuming t ∈ Hpws

(Γ ) for some s ∈ [− 12 , 1]. If the solution t ∈ Hpw
1
(Γ )
is sufficiently smooth we then conclude the optimal error estimate
3
t − th H −1/2 (Γ ) ≤ c h 2 |t|Hpw
1 (Γ )
when using piecewise constant basis functions.

Remark 12.1. When considering a direct boundary integral approach for the
Dirichlet boundary value problem of the Laplace equation, the solution
t ∈ H −1/2 (Γ ) is the exterior normal derivative t(x) = n(x) · ∇u(x) for x ∈ Γ .
When Γ = ∂Ω is the boundary of a domain with corners or edges, the ex-
terior normal vector n(x) is discontinuous and therefore the exterior normal
derivative t(x) is also discontinuous. Hence we have t ∈ H 2 −ε (Γ ) for any
d−1
arbitrary small ε > 0. When using globally continuous trial spaces, e.g. the
trial space Sh1 (Γ ) of piecewise linear basis functions, we can not apply the
error estimate (12.7) due to the global definition of Sh1 (Γ ), but we can apply
the global error estimate (12.6) for s < 12 (d − 1). Hence, when using higher
order basis functions they have to be discontinuous globally.
Up to now we have only considered error estimates in the energy norm

t − th H −1/2 (Γ ) . If the boundary mesh is globally quasi–uniform we can also
derive an error estimate in L2 (Γ ).
Lemma 12.2. For a globally quasi–uniform boundary mesh (10.2) let

th ∈ Sh0 (Γ ) be the unique solution of the Galerkin variational problem (12.5).
Let t ∈ Hpw s
(Γ ) for some s ∈ [0, 1] be satisfied. Then there holds the error
estimate
t − th L2 (Γ ) ≤ c hs |t|Hpw
s (Γ ) . (12.9)
Proof. For t ∈ L2 (Γ ) we define the L2 projection Qh t ∈ Sh0 (Γ ) as the unique

solution of the variational problem (10.4). By applying the triangle inequality
twice, and by using the global inverse inequality (see Lemma 10.10) we obtain
t − th L2 (Γ ) ≤ t − Qh tL2 (Γ ) + Qh t − th L2 (Γ )
≤ t − Qh tL2 (Γ ) + c h−1/2 Qh t − th H −1/2 (Γ )
0 1
≤ t − Qh tL2 (Γ ) + c h−1/2 t − Qh tH −1/2 (Γ ) + t − th H −1/2 (Γ ) .
The assertion now follows from the error estimates (10.6), (10.8), and (12.8).

After the approximate solution th ∈ Sh0 (Γ ) is determined as the unique
solution of the Galerkin variational problem (12.5), we obtain from (12.2) an
approximate solution of the Dirichlet boundary value problem (12.1), i.e. for
∈ Ω we have
x

∗ int U ∗ (
(
u x) = U (x, y)th (y)dsy − γ1,y x, y)g(y)dsy . (12.10)
Γ Γ
For the related error we first obtain

|u(
x) − u
(
x)| = U (∗
x, y)[t(y) − th (y)]dsy ∈ Ω.
for x

Γ
∈ Ω and y ∈ Γ the fundamental solution U ∗ (

Since for x x, y) is a C ∞ function,
i.e. infinitely often differentiable, we conclude U ∗ (
x, ·) ∈ H −σ (Γ ) for any
σ ∈ R. Hence we obtain
|u(
x) − u x)| ≤ U ∗ (
( x, ·)H −σ (Γ ) t − th H σ (Γ ) . (12.11)
To derive an almost optimal error estimate for |u( x) − u

( ∈ Ω, we need
x)|, x
to have an error estimate for t − th H σ (Γ ) where σ ∈ R is minimal.
Theorem 12.3 (Aubin–Nitsche Trick [82]). For some s ∈ [− 12 , 1] let

t ∈ Hpws
(Γ ) be the unique solution of the boundary integral equation (7.8),
and let th ∈ Sh0 (Γ ) be the unique solution of the Galerkin variational problem
(12.5). Assume that the single layer potential V : H −1−σ (Γ ) → H −σ (Γ ) is
continuous and bijective for some −2 ≤ σ ≤ − 12 . Then there holds the error
estimate
t − th H σ (Γ ) ≤ c hs−σ |t|Hpw
s (Γ ) . (12.12)
Proof. For σ < − 12 we have, by using the duality of Sobolev norms,
t − th , vΓ
t − th H σ (Γ ) = sup .
0=v∈H −σ (Γ ) vH −σ (Γ )
By assumption, the single layer potential V : H −1−σ (Γ ) → H −σ (Γ ) is bijec-

tive. Hence, for any v ∈ H −σ (Γ ) there exists a unique w ∈ H −1−σ (Γ ) such
that v = V w is satisfied. Therefore, by applying the Galerkin orthogonality
V (t − th ), τh Γ = 0 for all τh ∈ Sh0 (Γ )
we obtain
t − th , V wΓ
t − th H σ (Γ ) = sup
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
V (t − th ), w − Qh wΓ
= sup .
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
Since the single layer potential V : H −1−σ (Γ ) → H −σ (Γ ) is assumed to be

bijective, i.e. V wH −σ (Γ ) ≥ c wH −1−σ (Γ ) , we further conclude
w − Qh wH −1/2 (Γ )
t − th H σ (Γ ) ≤
c t − th H −1/2 (Γ ) sup .
0=w∈H −1−σ (Γ ) wH −1−σ (Γ )
When considering −1 − σ ≤ 1, i.e. σ ≥ −2, we obtain from the error estimate

(10.8)
w − Qh wH −1/2 (Γ )
sup ≤ c h−1/2−σ
0=w∈H −1−σ (Γ ) wH −1−σ (Γ )
and therefore
t − th H σ (Γ ) ≤ ĉ h−1/2−σ t − th H −1/2 (Γ ) .
The assertion now follows from the error estimate (12.8).

If the solution t of the boundary integral equation (7.8) is sufficiently
1
regular, i.e. t ∈ Hpw (Γ ), we obtain from (12.12) the optimal error estimate
t − th H −2 (Γ ) ≤ c h3 |t|Hpw
1 (Γ ) .
Moreover, when using (12.11) this gives the pointwise error estimate
|u(
x) − u x)| ≤ c h3 U ∗ (
( x, ·)H 2 (Γ ) |t|Hpw
1 (Γ ) .
To obtain a global error estimate for u − u

H 1 (Ω) we first consider the trace
∈ H 1 (Ω) which was defined via the represen-
g of the approximate solution u
tation formula (12.10). For x ∈ Γ we have
1
g(x) := (V th )(x) + g(x) − (Kg)(x).
2
On the other hand, the boundary integral equation (7.8) gives
1
g(x) = (V t)(x) + g(x) − (Kg)(x) for x ∈ Γ.
2
Hence, by subtraction we get
g(x) − g(x) = (V [t − th ])(x) for x ∈ Γ.
Theorem 12.4. Let u ∈ H 1 (Ω) be the weak solution of the Dirichlet bound-
ary value problem (12.1), and let u ∈ H 1 (Ω) be the approximate solution as
defined via the representation formula (12.10). Then there holds the global
error estimate
u − uH 1 (Ω) ≤ c t − th H −1/2 (Γ ) . (12.13)
Proof. The weak solution u = u0 + Eg ∈ H 1 (Ω) of the Dirichlet boundary

value problem (12.1) is given as the unique solution of the variational problem
a(u0 + Eg, v) = 0 for all v ∈ H01 (Ω).
Correspondingly, the approximate solution u =u g ∈ H 1 (Ω) as defined

0 + E
via the representation formula (12.10) satisfies
g , v) = 0 for all v ∈ H01 (Ω).

u0 + E
a(
Hence we have
a(u0 − u
0 , v) = a(E(
g − g), v) for all v ∈ H01 (Ω).
By choosing v := u0 − u 0 ∈ H01 (Ω) we obtain, by using the H01 (Ω) ellipticity

of the bilinear form a(·, ·),
cA 0 2H 1 (Ω) ≤ a(u0 − u
1 u0 − u 0 , u0 − u
0 ) = a(E(
g − g), u0 − u
0 )
2 E(
≤ cA g − g)H 1 (Ω) u0 − u
0 H 1 (Ω) ,
and therefore
cA
2
u0 − u
0 H 1 (Ω) ≤ E(
g − g)H 1 (Ω) .
cA
1
By applying the triangle inequality, and the inverse trace theorem, we conclude
u − u
H 1 (Ω) ≤ u0 − u
0 H 1 (Ω) + E(g − g)H 1 (Ω)

cA
≤ 1 + 2A E( g − g)H 1 (Ω)
c1

cA
≤ cIT 1 + 2A g − gH 1/2 (Γ )
c1

cA
≤ cIT 1 + 2A cV2 t − th H −1/2 (Γ )
c1
and therefore the claimed error estimate.
1
If t ∈ Hpw (Γ ), i.e. u ∈ H 5/2 (Ω), is satisfied, then from (12.13) we obtain
for the approximate solution u the error estimate in the energy space H 1 ,
H 1 (Ω) ≤ c h3/2 |t|Hpw

u − u 1 (Γ ) . (12.14)
Remark 12.5. When using lowest order, i.e. piecewise constant boundary el-
ements, and when assuming u ∈ H 5/2 (Ω) for the solution of the Dirichlet
boundary value problem, then the approximate solution u as defined via the
representation formula (12.10) converges with a convergence rate of 1.5. In
contrast, a lowest order, i.e. piecewise linear finite element solution converges
with a convergence rate of 1.0, see the error estimate (11.10), where we have
to assume u ∈ H 2 (Ω) only.
When inserting (12.4) into the Galerkin variational formulation (12.5), and
when choosing τh = ϕ0 as a test function, this gives

N
1
tk V ϕ0k , ϕ0 Γ = ( I + K)g, ϕ0 Γ for = 1, . . . , N.
2
k=1
By using
1
Vh [, k] = V ϕ0k , ϕ0 Γ ,
f = ( I + K)g, ϕ0 Γ
2
for k, = 1, . . . , N the variational problem is equivalent to a linear system of
algebraic equations,
Vh t = f (12.15)
where the stiffness matrix Vh is symmetric and positive definite, see Sec. 8.1.
Lemma 12.6. Let the boundary mesh be globally quasi–uniform. Then, for
all w ∈ RN ↔ wh ∈ Sh0 (Γ ) there hold the spectral equivalence inequalities
c1 hd w22 ≤ (Vh w, w) ≤ c2 hd−1 w22
with some positive constants ci , i = 1, 2.
Proof. The upper estimate follows from
(Vh w, w) = V wh , wh Γ ≤ cV2 wh 2H −1/2 (Γ )

N
≤ cV2 wh 2L2 (Γ ) = cV2 w 2 ∆
=1
by using ∆ = hd−1
.
To prove the lower estimate we consider an arbitrary but fixed w ∈ RN ↔
wh ∈ Sh0 (Γ ). The L2 projection QB
h wh ∈ Sh (Γ ) onto the space of local bubble
B
functions is defined as the unique solution of the variational problem (10.17),
h wh , τh L2 (Γ ) = wh , τh L2 (Γ )

QB for all τh ∈ Sh0 (Γ ).
Since the single layer potential V is H −1/2 (Γ ) elliptic, we first obtain
(Vh w, w) = V wh , wh Γ ≥ cV1 wh 2H −1/2 (Γ ) .
By using duality, the inverse inequality (10.16) in ShB (Γ ), and the stability
estimate (10.18) we further get
wh , vΓ wh , QB

h wh L2 (Γ )
wh H −1/2 (Γ ) = sup ≥
0=v∈H 1/2 (Γ ) vH 1/2 (Γ ) Qh wh H 1/2 (Γ )
B
wh , wh L2 (Γ ) wh 2L2 (Γ )

= = ≥ c h1/2 wh L2 (Γ )
QB
h w H 1/2 (Γ ) QBw
h h H 1/2 (Γ )
and therefore

N
(Vh w, w) ≥ c h wh 2L2 (Γ ) = c h w 2 ∆
=1
which finally gives the lower estimate.

To bound the spectral condition of the stiffness matrix Vh of the single
layer potential V we now obtain the estimate
κ2 (Vh ) ≤ c h−1 .
As an example we consider a uniform boundary mesh of the boundary Γ = ∂Ω

where Ω = (0, 0.5)2 is a square. By N we denote the number of boundary
L N λmin (Vh ) λmax (Vh ) κ2 (Vh )

2 16 2.04 –3 4.92 –2 24.14
3 32 5.14 –4 2.46 –2 47.86
4 64 1.29 –4 1.23 –2 95.64
5 128 3.22 –5 6.15 –3 191.01
6 256 8.06 –6 3.07 –3 381.32
7 512 2.02 –6 1.54 –3 760.73
8 1024 5.07 –7 7.68 –4 1516.02
theory: O(h2 ) O(h) O(h−1 )
Table 12.1. Spectral condition number of the stiffness matrix Vh .
elements τ while h = O(N −1 ) is the mesh size. In Table 12.1 there are given
the minimal and maximal eigenvalues of the stiffness matrix Vh as well as the
resulting spectral condition number κ2 (Vh ).
The computation of the load vector f requires the evaluation of

1
f = ( I + K)g(x)dsx for = 1, . . . , N
2
τ
where the double layer potential K is applied to the given Dirichlet data g. If
g is replaced by some piecewise linear approximation gh ∈ Sh1 (Γ ), we obtain

M
1
f = gi ( I + K)ϕ1i , ϕ0 Γ ,
i=1
2
in particular
1
f = ( Mh + Kh )g
2
where
Mh [, i] = ϕ1i , ϕ0 Γ , Kh [, i] = Kϕ1i , ϕ0 Γ
for i = 1, . . . , M and = 1, . . . , N . Hence we have to find the solution vector

t ∈ RN of the linear system
1
Vh
t = ( Mh + Kh )g.
2
The associated approximate solution

t ∈ Sh0 (Γ ) is the unique solution of the
perturbed variational problem
1
V
th , τh Γ = ( I + K)gh , τh Γ for all τh ∈ Sh0 (Γ ). (12.16)
2
When applying Theorem 8.2 (Strang Lemma) we then obtain the error esti-
mate

1
t −
th H −1/2 (Γ ) ≤ V cV2 inf t − τh H −1/2 (Γ ) + cW
2 g − gh H 1/2 (Γ )
c1 0 (Γ )
τh ∈Sh
1
≤ c1 hs+ 2 |t|Hpw
s (Γ ) + c2 g − gh 1/2
H (Γ ) (12.17)
when assuming t ∈ Hpw s

(Γ ) for some s ∈ [− 12 , 1]. It remains to estimate
the error of the approximation gh . When considering the piecewise linear
interpolation

M
gh (x) = Ih g(x) = g(xi )ϕ1i (x) ∈ Sh1 (Γ )
i=1
there holds the error estimate (10.11),
g − Ih gH σ (Γ ) ≤ c hs−σ |g|H s (Γ ) (12.18)
when assuming g ∈ H s (Γ ) for some s ∈ ( d−1 2 , 2] and 0 ≤ σ ≤ min{1, s}.

Instead of an interpolation we may also consider the piecewise linear L2 pro-
jection Qh g ∈ Sh1 (Γ ) as the unique solution of the variational problem
Qh g, vh L2 (Γ ) = g, vh L2 (Γ ) for all vh ∈ Sh1 (Γ ).
In this case a similar error estimate as in (10.8) holds, i.e.
g − Qh gH σ (Γ ) ≤ c hs−σ |g|H s (Γ ) (12.19)
when assuming g ∈ H s (Γ ) for some 0 ≤ s ≤ 2 and −1 ≤ σ ≤ min{1, s}. In

both cases we find from (12.17) the error estimate
1
σ− 12
t −
th H −1/2 (Γ ) ≤ c1 hs+ 2 |t|Hpw
s (Γ ) + c2 h |g|H σ (Γ )
when assuming t ∈ Hpw s

(Γ ) for some s ∈ [− 12 , 1], and g ∈ H σ (Γ ) for some
1
σ ∈ [ 2 , 2] in the case of the L2 projection, and σ ∈ ( d−1 2 , 2] in the case of
1
interpolation. In particular, when assuming t ∈ Hpw (Γ ) and g ∈ H 2 (Γ ) we
! "
t −
th H −1/2 (Γ ) ≤ c h3/2 |t|Hpw
1 (Γ ) + |g|H 2 (Γ ) .
For the error estimate t − th H −1/2 (Γ ) in the energy space it is not impor-
tant whether the given data are replaced by some interpolation or by some
projection. However, this changes when considering error estimates in lower
Sobolev spaces, as they are used to obtain error estimates for the approximate
solution as defined via the representation formula (12.10).
Theorem 12.7. Let t ∈ Hpw s

(Γ ) for some s ∈ [− 12 , 1] be the unique solution
of the boundary integral equation (7.8), and let th ∈ Sh0 (Γ ) be the unique
solution of the perturbed variational problem (12.16). Assume that the single
layer potential V : H −1−σ (Γ ) → H −σ (Γ ) is bounded and bijective for some
σ ∈ [−2, − 12 ], and let the double layer potential 12 I +K : H 1+σ (Γ ) → H 1+σ (Γ )
be bounded. Moreover, let g ∈ H (Γ ) for some ∈ (1, 2]. Then there holds the
error estimate
t −
th H σ (Γ ) ≤ c1 hs−σ |t|Hpw
s (Γ ) + c2 h
−σ−1
|g|H (Γ )
where σ ≥ −1 in the case of interpolation gh = Ih g, and σ ≥ −2 in the case

of the L2 projection gh = Qh g.
Proof. As in the proof of Theorem 12.3 we have for σ ∈ [−2, − 12 )
V (t −
th ), wΓ
t −
th H σ (Γ ) = sup .
0=w∈H −1−σ (Γ ) V w H −σ (Γ )
By subtracting the perturbed variational formulation (12.16) from the exact

formulation (12.3) we conclude the equality
1
V (t −
th ), τh Γ = ( I + K)(g − gh ), τh Γ for all τh ∈ Sh0 (Γ ).
2
Hence we obtain
V (t −
th ), w − Qh wΓ
t −
th H σ (Γ ) ≤ sup
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
( 12 I + K)(g − gh ), Qh wΓ
+ sup .
0=w∈H −1−σ (Γ ) V wH −σ (Γ )
As in the proof of Theorem 12.3 we get
V (t −
th ), w − Qh wΓ
sup ≤ c hs−σ |t|Hpw
s (Γ )
0=w∈H −1−σ (Γ ) V w H −σ (Γ )
when assuming σ ≥ −2. To estimate the second term we have

1 1
|( I + K)(g − gh ), Qh wΓ | = |( I + K)(g − gh ), wΓ |
2 2
1
+|( I + K)(g − gh ), w − Qh wΓ |
2
≤ c1 g − gh H 1+σ (Γ ) wH −1−σ (Γ ) + c2 g − gh H 1 (Γ ) w − Qh wH −1 (Γ )
≤ c1 g − gh H 1+σ (Γ ) wH −1−σ (Γ ) + c2 h−1 |g|H (Γ ) h−σ wH −1−σ (Γ )
and therefore
t −
th H σ (Γ ) ≤ c1 hs−σ |t|Hpw
s (Γ ) + c2 h
−σ−1
|g|H (Γ ) + c3 g − gh H 1+σ (Γ ) .
Finally, for ∈ (1, 2] we have
g − gh H 1+σ (Γ ) ≤ c h−σ−1 |g|H (Γ )
for some 1 + σ ≥ 0 in the case of interpolation gh = Ih g, and 1 + σ ≥ −1 in

the case of the L2 projection gh = Qh g, see the error estimates (12.18) and
(12.19), respectively. This finishes the proof.
1
When assuming t ∈ Hpw (Γ ) and g ∈ H 2 (Γ ) we therefore obtain the error
estimate ! "
t −
th H σ (Γ ) ≤ c h1−σ |t|Hpw 1 (Γ ) + |g|H 2 (Γ ) .
In the case of an interpolation gh = Ih g we obtain by choosing σ = −1 the
optimal error estimate
! "
t −
th H −1 (Γ ) ≤ c h2 |t|Hpw
1 (Γ ) + |g|H 2 (Γ ) ,
while in the case of the L2 projection gh = Qh g we obtain by choosing σ = −2

the improved estimate
! "
t − th H −2 (Γ ) ≤ c h3 |t|Hpw
1 (Γ ) + |g|H 2 (Γ ) .
An approximate solution of the Dirichlet boundary value problem (12.1) is

then given via the representation formula

û(
x) = x, y)
U ∗ ( int U ∗ (
th (y)dsy − γ1,y ∈ Ω (12.20)
x, y)gh (y)dsy for x
Γ Γ
satisfying the error estimate

! "
|u( x)| ≤ c h2
x) − û( 1 (Γ ) + |g|H 2 (Γ )
|t|Hpw (12.21)
in the case of an interpolation gh = Ih g, and

! "
|u( x)| ≤ c h3 |t|Hpw
x) − û( 1 (Γ ) + |g|H 2 (Γ ) (12.22)
in the case of the L2 projection gh = Qh g.

To check the theoretical convergence results, and to compare boundary and
finite element methods we now consider a simple Dirichlet boundary value
problem (12.1) where the domain Ω = (0, 0.5)2 is a square, and where the
Dirichlet data are such that the solution is given by (11.14),
1
u(x) = − log |x − x∗ |, x∗ = (−0.1, −0.1) .
2
The boundary Γ = ∂Ω is discretized into N boundary elements τ of mesh
size h. The numerical results as given in Table 12.2 confirm the theoretical
FEM interpolation L2 projection

N λ − λh L2 (Γ ) t − th L2 (Γ ) eoc t − th L2 (Γ ) eoc
8 9.350 –1 8.774 –1 8.525 –1
16 5.662 –1 5.217 –1 0.75 5.180 –1 0.72
32 2.970 –1 2.745 –1 0.93 2.751 –1 0.91
64 1.457 –1 1.384 –1 0.99 1.387 –1 0.99
128 7.093 –2 6.897 –2 1.00 6.905 –2 1.01
256 3.482 –2 3.433 –2 1.01 3.434 –2 1.01
512 1.723 –2 1.711 –2 1.00 1.711 –2 1.01
1024 8.569 –3 8.539 –3 1.00 8.539 –3 1.00
2048 4.265 –3 1.00 4.265 –3 1.00
4096 2.131 –3 1.00 2.131 –3 1.00
theory: 1 1
Table 12.2. Error and order of convergence, Dirichlet boundary value problem.
estimate (12.9) which was obtained for the error t − th L2 (Γ ) . In addition we

also give the errors λ − λh L2 (Γ ) of the Lagrange multipliers as obtained by
a mixed finite element method, see also Table 11.3.
Finally we also consider the error of the approximate solution (12.20).
The results in the interior point x̂ = (1/7, 2/7) as listed in Table 12.3 again
confirm the theoretical error estimates (12.21) and (12.22).
interpolation L2 projection
N |u(x̂) − û(x̂)| eoc |u(x̂) − û(x̂)| eoc
8 2.752 –2 6.818 –3
16 5.463 –3 2.33 5.233 –4 3.70
32 1.291 –3 2.08 6.197 –5 3.08
64 3.147 –4 2.04 6.753 –6 3.20
128 7.780 –5 2.02 7.587 –7 3.15
256 1.935 –5 2.01 8.878 –8 3.10
512 4.827 –6 2.00 1.070 –8 3.05
1024 1.205 –6 2.00 1.312 –9 3.03
2048 3.012 –7 2.00 1.620 –10 3.02
4096 7.528 –8 2.00 1.921 –11 3.08
theory: 2 3
Table 12.3. Error and order of convergence, interior point.
Next we consider the Neumann boundary value problem (7.16),

Lu(x) = 0 for x ∈ Ω, γ1int u(x) = g(x) for x ∈ Γ,
where we have to assume the solvability condition, see (1.17),

g(x)dsx = 0.
Γ
The solution is given via the representation formula (6.1) for x ∈ Ω,

u(
x) = U ∗ ( int U ∗ (
x, y)g(y)dsy − γ1,y x, y)γ0int u(y)ds. (12.23)
Γ Γ
Hence we have to find the yet unknown Dirichlet datum γ0int u ∈ H 1/2 (Γ ) as
the unique solution of the stabilized variational problem (7.24) such that
1
Dγ0int u, vΓ + γ0int u, weq Γ v, weq Γ = ( I − K )g, vΓ (12.24)
2
is satisfied for all v ∈ H 1/2 (Γ ). Let
Sh1 (Γ ) = span{ϕ1i }M
i=1 ⊂ H
1/2
(Γ )
be the trial space of piecewise linear and continuous basis functions ϕ1i . By
using
M
uh (x) = ui ϕ1i (x) ∈ Sh1 (Γ )
i=1
the Galerkin variational formulation of (12.24) reads to find uh ∈ Sh1 (Γ ) such

that
1
Duh , vh Γ + uh , weq Γ vh , weq Γ = ( I − K )g, vh Γ (12.25)
2
is satisfied for all vh ∈ Sh1 (Γ ). Since all assumptions of Theorem 8.1 (Cea’s
Lemma) are satisfied, there exists a unique solution uh ∈ Sh1 (Γ ) of the
Galerkin variational problem (12.25) satisfying both the stability estimate
uh H 1/2 (Γ ) ≤ c gH −1/2 (Γ )
γ0int u − uh H 1/2 (Γ ) ≤ c inf

1 (Γ )
γ0int u − vh H 1/2 (Γ ) .
vh ∈Sh
Due to
1
ker D = ker( I + K) = span{1} ⊂ Sh1 (Γ )
2
we find from the solvability assumption (1.17) the orthogonality relation
uh , weq Γ = 0
1/2
and therefore uh ∈ H∗ (Γ ).
By using the approximation property (10.14) we further obtain the error
estimate
γ0int u − uh H 1/2 (Γ ) ≤ c hs−1/2 |γ0int u|H s (Γ ) (12.26)
when assuming γ0int u ∈ H s (Γ ) for some s ∈ [ 12 , 2]. When γ0int u ∈ H 2 (Γ ) is

sufficiently smooth, we then conclude the error estimate
γ0int u − uh H 1/2 (Γ ) ≤ c h3/2 |γ0int u|H 2 (Γ ) .
Theorem 12.8 (Aubin–Nitsche Trick [82]). For some s ∈ [ 12 , 2] let

γ0int u ∈ H s (Γ ) be the unique solution of the variational problem (7.24), and
let uh ∈ Sh1 (Γ ) be the unique solution of the Galerkin variational problem
(12.25). Assume that the stabilized hypersingular boundary integral operator
D : H 1−σ (Γ ) → H −σ (Γ ) is bounded and bijective for some σ ∈ [−1, 1 ]. Then
2
γ0int u − uh H σ (Γ ) ≤ c hs−σ |γ0int u|H s (Γ ) .
Proof. When considering σ ≤ 0 we have, by using the duality of Sobolev

norms,
γ0int u − uh , wΓ
γ0int u − uh H σ (Γ ) = sup .
0=w∈H −σ (Γ ) wH −σ (Γ )
By assumption, the hypersingular integral operator D : H 1−σ (Γ ) → H −σ (Γ )

is bijective. Hence, for any w ∈ H (Γ ) there exists a unique v ∈ H 1−σ (Γ )
−σ
such that w = Dv is satisfied. Therefore, by applying the Galerkin orthogo-

nality
int u − uh ), vh Γ = 0 for all vh ∈ S 1 (Γ )
D(γ 0 h
we obtain
Γ
γ0int u − uh , Dv
γ0int u − uh H σ (Γ ) = sup
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
int u − uh ), v − Q vΓ
D(γ
1/2
0 h
= sup
0=v∈H 1−σ (Γ )
Dv H −σ (Γ )
1/2
where Qh : H 1/2 (Γ ) → Sh1 (Γ ) is the projection which is defined via the
variational formulation
1/2
Qh v, zh H 1/2 (Γ ) = v, zh H 1/2 (Γ ) for all zh ∈ Sh1 (Γ ).
Note that there holds the error estimate (10.13), i.e. for s ∈ [1, 2] we have
1/2
v − Qh vH 1/2 (Γ ) ≤ c hs−1/2 |v|H s (Γ ) .
Since the stabilized hypersingular boundary integral operator D : H −σ (Γ ) →

H 1−σ
(Γ ) is assumed to be bijective, i.e. DvH −σ (Γ ) ≥ c vH 1−σ (Γ ) , we
further conclude
1/2
1 int v − Qh vH 1/2 (Γ )
γ0int u − uh H σ (Γ ) ≤ γ0 u − uh H 1/2 (Γ ) sup .
c 0=v∈H 1−σ (Γ ) vH 1−σ (Γ )
When considering 1 − σ ≤ 2, i.e. σ ≥ −1, we further obtain
1/2
v − Qh vH 1/2 (Γ )
sup ≤ c h1/2−σ
0=v∈H 1−σ (Γ ) vH 1−σ (Γ )
and therefore
γ0int u − uh H σ (Γ ) ≤ c h1/2−σ γ0int u − uh H 1/2 (Γ ) .
Now, the assertion follows from (12.26). Finally, for σ ∈ (0, 12 ) we can apply
an interpolation argument.
When assuming γ0int u ∈ H 2 (Γ ) we then find the optimal error estimate
γ0int u − uh H −1 (Γ ) ≤ c h3 |γ0int u|H 2 (Γ ) .
Inserting the unique solution uh into the representation formula (12.23) this
defines an approximate solution of the Neumann boundary value problem
(7.16),

∗ int U ∗ (
(
u x) = x, y)g(y)dsy − γ1,y
U ( ∈ Ω,
x, y)uh (y)dsy for x
Γ Γ
which satisfies, when assuming γ0int u ∈ H 2 (Γ ), the error estimate

|u(
x) − u x)| = γ1int U ∗ (
( x, y)[γ0int u(y) − uh (y)]dsy

Γ
≤ γ1int U ∗ (
x, ·)H 1 (Γ ) γ0int u − uh H −1 (Γ )
≤ c h3 γ1int U ∗ (
x, ·)H 1 (Γ ) |γ0int u|H 2 (Γ ) .
The Galerkin variational problem (12.25) is equivalent to a linear system of
algebraic equations,
(Dh + α a a )u = f (12.27)
where
1
Dh [j, i] = Dϕ1i , ϕ1j Γ , aj = ϕ1j , weq Γ ,
fj = ( I − K )g, ϕ1j Γ
2
h := Dh + α a a is
for i, j = 1, . . . , M . The stabilized stiffness matrix D
symmetric and positive definite, see also the general discussion in Sec. 8.1.
Lemma 12.9. Let the boundary mesh be globally quasi–uniform. Then, for
all v ∈ RM ↔ vh ∈ Sh1 (Γ ) there hold the spectral equivalence inequalities
h v, v) ≤ c2 hd−2 v22
c1 hd−1 v22 ≤ (D
with some positive constants ci , i = 1, 2.

Proof. Since the stabilized hypersingular boundary integral operator
D : H 1/2 (Γ ) → H −1/2 (Γ ) is bounded and H 1/2 (Γ ) elliptic, we first obtain
the spectral equivalence inequalities
2 2
1 vh H 1/2 (Γ ) ≤ (Dh v, v) ≤ c2 vh H 1/2 (Γ )
cD D
for all vh ∈ Sh1 (Γ ).
By using the inverse inequality in Sh1 (Γ ), and by applying Lemma 10.5 we

further have

N
vh 2H 1/2 (Γ ) ≤ cI h−1 vh 2L2 (Γ ) = c h−1 ∆ vk2
=1 k∈J( )
and with ∆ = hd−1

the upper estimate follows. The lower estimate finally
results from

N
vh 2H 1/2 (Γ ) ≥ vh 2L2 (Γ ) ≥ c ∆ vk2 .

=1 k∈J( )
To bound the spectral condition number of the stiffness matrix D h of

the stabilized hypersingular boundary integral operator we now obtain the
estimate
h ) ≤ c h−1 .
κ2 (D
For the computation of the load vector f again we can introduce, as for the
Dirichlet boundary value problem, a projection of the given Neumann data
g ∈ H −1/2 (Γ ) onto the space of piecewise constant basis functions. When
assuming g ∈ L2 (Γ ) the L2 projection Qh g ∈ Sh0 (Γ ) is defined as the unique
Qh g, τh L2 (Γ ) = g, τh L2 (Γ ) for all τh ∈ Sh0 (Γ ).
Note that there holds the error estimate
g − Qh gH σ (Γ ) ≤ c hs−σ |g|Hpw
s (Γ )
when assuming g ∈ Hpw s

(Γ ) for some σ ∈ [−1, 0], if s ∈ [0, 1] is satisfied. By
using gh = Qh g the perturbed variational problem is to find u h ∈ Sh1 (Γ ) such
that
1
D uh , weq Γ vh , weq Γ = ( I − K )gh , vh Γ
uh , vh Γ + α (12.28)
2
is satisfied for all vh ∈ Sh1 (Γ ). This is equivalent to a linear system of equations,
1
(Dh + α a a )
u = ( Mh − Kh )g .
2
If R ⊂ Sh0 (Γ ) is satisfied, then we have
Qh g, vk L2 (Γ ) = g, vk L2 (Γ ) = 0 for all vk ∈ R
and therefore the solvability condition

Qh g(x)vk (x)dsx = 0 for all vk ∈ R
Γ
1/2
is satisfied, i.e. uh ∈ H∗ (Γ ).
The application of Theorem 8.2 (Strang Lemma) gives, when assuming
γ0int u ∈ H s (Γ ) and g ∈ Hpw

(Γ ) for some s ∈ [ 12 , 2] and ∈ [0, 1], respectively,
the error estimate

cD
2 cW
2
γ0int u − u
h H 1/2 (Γ ) ≤
inf1 γ0int u − vh H 1/2 (Γ ) +
g − gh H −1/2 (Γ )
cD v ∈Sh (Γ ) cD
1 h 1
1
≤ c1 hs−1/2 |γ0int u|H s (Γ ) + c2 h+ 2 |g|H (Γ ) . (12.29)
In particular, when assuming γ0int u ∈ H 2 (Γ ) and g ∈ Hpw 1

(Γ ) we then obtain
the error estimate
! "
h H 1/2 (Γ ) ≤ c h3/2 |γ0int u|H 2 (Γ ) + |g|H 1 (Γ ) .
γ0int u − u
Theorem 12.10. For some s ∈ [ 12 , 2] let γ0int u ∈ H s (Γ ) be the unique solu-

tion of the variational problem (7.24), and let u h ∈ Sh1 (Γ ) be the unique solu-

tion of the perturbed Galerkin variational problem (12.28) where g ∈ Hpw (Γ )
for some ∈ [− 12 , 1]. Assume that the stabilized hypersingular boundary in-
tegral operator D : H 1−σ (Γ ) → H −σ (Γ ) is bounded and bijective for some
σ ∈ [−1, 12 ], and let the adjoint double layer potential 12 I − K : H −1+σ (Γ ) →
H −1+σ (Γ ) be bounded. Then there holds the error estimate
γ0int u − u
h H σ (Γ ) ≤ c1 hs−σ |γ0int u|H s (Γ ) + c2 h+1−σ |g|Hpw

(Γ )
where σ ∈ [0, 12 ].
Proof. As in the proof of Theorem 12.8 we have for some σ ∈ [−1, 12 ]
int u − u
D(γ h ), vΓ
0
γ0int u − u
h H σ (Γ ) = sup .
0=v∈H 1−σ (Γ ) H −σ (Γ )
Dv
By subtracting the perturbed variational formulation (12.28) from the exact

formulation (12.25) we conclude the equality
int u − u 1
D(γ0 h ), vh Γ = ( I − K )(g − gh ), vh Γ for all vh ∈ Sh1 (Γ ).
2
Therefore,
int u − u
D(γ
1/2
h ), v − Qh vΓ
0
γ0int u − u
h H σ (Γ ) ≤ sup
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
1/2
( 12 I − K )(g − gh ), Qh vΓ
+ sup
0=v∈H 1−σ (Γ ) H −σ (Γ )
Dv
As in the proof of Theorem 12.8 we have
int u − u
D(γ h ), v − Qh vΓ
1/2
0
sup ≤ c1 hs−σ |γ0int u|H s (Γ )
0=v∈H 1−σ (Γ )
DvH −σ (Γ )
for some σ ≥ −1. The second term can be estimated by

1 1/2 1
|( I − K )(g − gh ), Qh vΓ | ≤ |( I − K )(g − gh ), vΓ |
2 2
1 1/2
+|( I − K )(g − gh ), v − Qh vΓ |
2
1/2
≤ c2 g − gh H −1+σ (Γ ) vH 1−σ (Γ ) + c3 g − gh L2 (Γ ) v − Qh vL2 (Γ )
1−σ
≤ c2 g − gh H −1+σ (Γ ) vH 1−σ (Γ ) + c3 h |g|Hpw

(Γ ) h vH 1−σ (Γ )
and therefore we get
γ0int u − u
h H σ (Γ ) ≤ c1 hs−σ |γ0int u|H s (Γ ) + c2 h+1−σ |g|Hpw

(Γ )
+c3 g − gh H −1+σ (Γ ) .
Finally, for the L2 projection gh = Qh g there holds the error estimate
g − Qh gH −1+σ (Γ ) ≤ c h+1−σ |g|Hpw

(Γ )

for −1 + σ ≥ −1, i.e. σ ≥ 0, and when assuming g ∈ Hpw (Γ ) for some
∈ [0, 1].

When assuming γ0int u ∈ H 2 (Γ ) and g ∈ Hpw 1
(Γ ) we therefore obtain the
error estimate
! "
h L2 (Γ ) ≤ c h2 |γ0int u|H 2 (Γ ) + |g|Hpw
γ0int u − u 1 (Γ ) ,
and in the sequel we can compute the approximate solution

∗ int U ∗ (
û(
x) = U (x, y)gh (y)dsy − γ1,y x, y)uh (y)dsy for x ∈ Ω
Γ Γ
which satisfies the pointwise error estimate

! "
|u( x)| ≤ c h2 |γ0int u|H 2 (Γ ) + |g|Hpw
x) − û( 1 (Γ ) .
Hence, when using approximate boundary conditions, in particular when using

a piecewise constant L2 projection Qh g ∈ Sh0 (Γ ) of the given Neumann data
g this results in non–optimal error estimates, i.e. the order of convergence is
reduced in comparison to the case when using the exact boundary data.
Instead of the stabilized variational formulation (7.24) we may also con-
sider the modified variational problem (7.25). Then, the associated Galerkin
variational formulation is to find uh ∈ Sh1 (Γ ) such that
1
Duh , vh Γ + ᾱ uh , 1Γ vh , 1Γ = ( I − K )g, vh Γ (12.30)
2
is satisfied for all vh ∈ Sh1 (Γ ). From the solvability assumption (1.17) we
1/2
then find uh ∈ H∗∗ (Γ ). All error estimates for the approximate solution
1
uh ∈ Sh (Γ ) follow as in Theorem 12.8. The variational problem (12.30) is
equivalent to a linear system of algebraic equations,

Dh + ᾱ ā ā u = f ,
where
āj = ϕ1j , 1Γ for j = 1, . . . , M.
12.3 Mixed Boundary Conditions

The solution of the mixed boundary value problem (7.35),
Lu(x) = 0 for x ∈ Ω
γ1int u(x) = gN (x) for x ∈ ΓN
is given via the representation formula (7.36).

By gD ∈ H 1/2 (Γ ) and gN ∈ H −1/2 (Γ ) we denote some appropriate exten-
sions of the given data gD ∈ H 1/2 (ΓD ) and gN ∈ H −1/2 (ΓN ), respectively.
Then, the symmetric formulation of boundary integral equations reads to find
the yet unknown Cauchy data ( ) ∈ H
t, u 1/2 (ΓN ) as the unique
−1/2 (ΓD ) × H
solution of the variational problem (7.39) such that
a( ; τ, v) = F (τ, v)
t, u
−1/2 (ΓD ) × H
is satisfied for all (τ, v) ∈ H 1/2 (ΓN ). By using

ND

th (x) =
tk ϕ0k (x) ∈ Sh0 (ΓD ),
k=1

MN
h (x) =
u u 1/2 (ΓN )
i ϕ1i (x) ∈ Sh1 (ΓN ) ∩ H
i=1
the approximate solution ( h ) is given as the unique solution of the

th , u
Galerkin variational formulation such that
a( h ; τh , vh ) = F (τh , vh )
th , u (12.31)
1/2 (ΓN ). Based on Lemma

is satisfied for all (τh , vh ) ∈ Sh0 (ΓD ) × Sh1 (ΓN ) ∩ H
7.4 all assumptions of Theorem 8.1 (Cea’s Lemma) are satisfied. Therefore we
have the unique solvability of the Galerkin variational problem (12.31) as well
as the a priori error estimate
h 2H 1/2 (Γ ) +
u−u t−
th 2H −1/2 (Γ ) (12.32)
+
≤c inf u − vh 2H 1/2 (Γ ) +
inf
0 (Γ )

t − τh 2H −1/2 (Γ ) .
vh ∈Sh N
-1/2 (ΓN )
1 (Γ )∩H τh ∈Sh D
Assume that the solution u of the mixed boundary value problem (7.35) sat-
isfies u ∈ H s (Ω) for some s > 32 . Applying the trace theorem we obtain
γ0int u ∈ H s−1/2 (Γ ) as well as γ1int u ∈ Hpw
s−3/2
(Γ ). When assuming that the
extensions gD ∈ H s−1/2
(Γ ) and gN ∈ Hpw
s−3/2
(Γ ) are sufficiently regular, and
using the approximation properties of the trial spaces Sh0 (ΓD ) and Sh1 (ΓN ) we
obtain from (12.32) the error estimate
h 2H 1/2 (Γ ) +
u−u t− u|2H σ1 (Γ ) + c2 h2σ2 +1 |
th 2H −1/2 (Γ ) ≤ c1 h2σ1 −1 | t|2Hpw
σ2
(Γ )
where

1 1 1 3
≤ σ1 ≤ min 2, s − , − ≤ σ2 ≤ min 1, s − .
2 2 2 2
In particular for σ1 = σ and σ2 = σ − 1 we have the error estimate

! "

u−u h 2H 1/2 (Γ ) +
t− u|2H σ (Γ ) + |
th 2H −1/2 (Γ ) ≤ c h2σ−1 | t|2Hpw
σ−1
(Γ )
1
for all 2 ≤ σ ≤ min{2, s − 12 }. If u ∈ H 5/2 (Ω) is sufficiently smooth, we then
have
! "
h 2H 1/2 (Γ ) +
u−u t−
th 2H −1/2 (Γ ) ≤ c h3 u|2H 2 (Γ ) + |
| t|2Hpw
1 (Γ ) ,
i.e. the order of convergence for approximate solutions of the mixed bound-
ary value problem corresponds to those of the pure Dirichlet or Neumann
boundary value problem. As for the pure Dirichlet or Neumann boundary

value problem we may also obtain error estimates in other Sobolev norms.
In particular, as in Theorem 12.10 (Aubin–Nitsche Trick) we obtain, when
assuming u ∈ H 2 (Γ ), the error estimate
2 31/2

u−u u|2H 2 (Γ ) + |
h L2 (Γ ) ≤ c h2 | t|2Hpw
1 (Γ ) . (12.33)
In analogy to Lemma 12.2 we finally have

2 31/2

t− u|2H 2 (Γ ) + |
th L2 (Γ ) ≤ c h | t|2Hpw
1 (Γ ) . (12.34)
The Galerkin variational formulation (12.31) is equivalent to a linear system

of algebraic equations,
& '& '
Vh −Kh
t f1

= (12.35)
Kh Dh
u f2
where the block matrices are given by
Vh [, k] = V ϕ0k , ϕ0 ΓD , Dh [j, i] = Dϕ1i , ϕ1j ΓN , Kh [, i] = Kϕ1i , ϕ0 ΓD
and where the load vectors are defined via

1 1
gD − V gN , ϕ0 ΓD , f2,j = ( I − K )
f1, = ( I + K) gD , ϕ1j ΓN
gN − D
2 2
for k, = 1, . . . , ND ; i, j = 1, . . . , MN .
The stiffness matrix in (12.35) is block skew–symmetric and positive defi-
nite, or symmetric and indefinite. From the H −1/2 (ΓD )–ellipticity of the single
layer potential V we conclude the positive definiteness of the Galerkin matrix
Vh . Hence we can solve the first equation in (12.35) for the first unknown t,
i.e.

t = Vh−1 Kh u + Vh−1 f 1 .
Inserting this into the second equation of (12.35) this gives the Schur comple-
ment system 0 1
Dh + Kh Vh−1 Kh u = f 2 − Kh Vh−1 f 1
where the Schur complement matrix
Sh = Dh + Kh Vh−1 Kh (12.36)
is symmetric and positive definite.

If we use appropriate approximations of the given boundary data to eval-
uate the load vectors in (12.35), we can apply Theorem 8.2 to obtain the
corresponding error estimates.
In what follows we will describe an alternative approach to solve the mixed

boundary value problem (7.35). For this we consider the variational formula-
tion (7.40) to find u∈H 1/2 (ΓN ) such that
S
u, vΓ = gN − S
gD , vΓN
1/2 (ΓN ). Recall that
is satisfied for all v ∈ H
1 1
S := D + ( I + K )V −1 ( I + K) : H 1/2 (Γ ) → H −1/2 (Γ )
2 2
is the Steklov–Poincaré operator. By using

MN
h (x) =
u u 1/2 (ΓN )
i ϕ1i (x) ∈ Sh1 (Γ ) ∩ H
i=1
we obtain an associated Galerkin variational formulation and the correspond-

= f , where
ing linear system of algebraic equations, Sh u
Sh [j, i] = Sϕ1i , ϕ1j ΓN , fj = gN − S

gD , ϕj ΓN
for i, j = 1, . . . , MN . However, due to the inverse single layer potential V −1 ,
the symmetric representation of the Steklov–Poincaré operator S does not
allow a direct computation of the stiffness matrix Sh . Hence we first have to
define a symmetric approximation of the continuous Steklov–Poincaré opera-
tors S.
Let v ∈ H 1/2 (Γ ) be some given function. Then, the application of the
Steklov–Poincaré operator by using the symmetric representation (6.43) reads
1 1 1
Sv = Dv + ( I + K )V −1 ( I + K)v = Dv + ( I + K )w
2 2 2
where w = V −1 ( 12 I + K)v ∈ H −1/2 (Γ ), i.e. w ∈ H −1/2 (Γ ) is the unique
1
V w, τ Γ = ( I + K)v, τ Γ for all τ ∈ H −1/2 (Γ ).
2
Let Sh0 (Γ ) = span{ϕ0k }N
k=1 ⊂ H
−1/2
(Γ ) be the conforming trial space of
piecewise constant basis functions, then the associated Galerkin variational
formulation is to find wh ∈ Sh0 (Γ ) such that
1
V wh , τh Γ = ( I + K)v, τh Γ (12.37)
2
is satisfied for all τh ∈ Sh0 (Γ ). By
= Dv + ( 1 I + K )wh
Sv (12.38)
2
we then define an approximation S of the Steklov–Poincaré operator S.
Lemma 12.11. The approximate Steklov–Poincareé operator S as defined in

(12.38) is bounded, i.e. S : H 1/2 (Γ ) → H −1/2 (Γ ) satisfying

H −1/2 (Γ ) ≤ cS2 vH 1/2 (Γ )
Sv for all v ∈ H 1/2 (Γ ).
Moreover, S is H
1/2 (ΓN )–elliptic,
vΓ ≥ cD v2 1/2
Sv, 1/2 (ΓN ),
for all v ∈ H
1 H (Γ )
and satisfies the error estimate

H −1/2 (Γ ) ≤ c
(S − S)v inf Sv − τh H −1/2 (Γ ) .
0 (Γ )
τh ∈Sh
Proof. For the solution wh ∈ Sh0 (Γ ) of the Galerkin variational problem

(12.37) we have, by using the H −1/2 (Γ )–ellipticity of the single layer potential
V,
1
cV1 wh 2H −1/2 (Γ ) ≤ V wh , wh Γ = ( I + K)v, wh Γ
2
1
≤ ( + cK 2 ) vH 1/2 (Γ ) wh H −1/2 (Γ )
2
and therefore
1 1
wh H −1/2 (Γ ) ≤ 2 ) vH 1/2 (Γ ) .
( + cK
cV1 2
Then we can conclude the boundedness of the approximate Steklov–Poincaré
operator S by
H −1/2 (Γ ) = Dv + ( 1 I + K )wh H −1/2 (Γ )

Sv
2
1
≤ cD
2 vH 1/2 (Γ ) + ( + c2 ) wh H −1/2 (Γ )
K
2

1 1 K 2
≤ c2 + V ( + c2 ) vH 1/2 (Γ ) .
D
c1 2
1/2 (ΓN ) ellipticity of S follows due to

The H
vΓ = Dv, vΓ + ( 1 I + K )wh , vΓ

Sv,
2
1
= Dv, vΓ + wh , ( I + K)vΓ
2
2
= Dv, vΓ + V wh , wh Γ ≥ cD 1 vH 1/2 (Γ )
from the H −1/2 (Γ ) ellipticity of the single layer potential V and from the
H 1/2 (ΓN ) ellipticity of the hypersingular boundary integral operator D.
By taking the difference
= ( 1 I + K )(w − wh )
(S − S)v
2
we finally obtain
H −1/2 (Γ ) = ( 1 I + K )(w − wh )H −1/2 (Γ )

(S − S)v
2
1
≤ ( + cK 2 )w − wh H −1/2 (Γ ) ,
2
and by using the error estimate (12.6) for the Dirichlet boundary value prob-
lem we complete the proof.
We now consider a modified Galerkin variational formulation to find an
1/2 (ΓN ) such that
approximate solution ûh ∈ Sh1 (Γ ) ∩ H
Sûh , ϕ1j ΓN = gN − S

gD , ϕ1 Γ
j N (12.39)
is satisfied for all j = 1, . . . , MN . By using Theorem 8.3 (Strang Lemma) we

conclude that the perturbed variational problem (12.39) is unique solvable,
and there holds the error estimate

u − ûh H 1/2 (Γ )
+
≤c inf
u − vh H 1/2 (Γ ) + inf0 Su − τh H −1/2 (Γ ) .
-1/2 (ΓN )
1 (Γ )∩H
vh ∈Sh τh ∈Sh (Γ )
When assuming u ∈ H 2 (Γ ) and Su ∈ Hpw 1

(Γ ) we then obtain the error
estimate
! "
u − ûh H 1/2 (Γ ) ≤ c h3/2 uH 2 (Γ ) + SuHpw
1 (Γ ) .
As in Theorem 12.10 we also find the error estimate

! "
u − ûh L2 (Γ ) ≤ c h2 uH 2 (Γ ) + SuHpw
1 (Γ ) . (12.40)
If the complete Dirichlet datum uh := ûh + gD is known we can solve a

Dirichlet boundary value problem to find the approximate Neumann datum
th ∈ Sh0 (Γ ) satisfying the error estimate
! "
t − th L2 (Γ ) ≤ c h uH 2 (Γ ) + SuHpw
1 (Γ ) . (12.41)
The Galerkin variational formulation (12.39) is equivalent to a linear system

of algebraic equations, Sh û = f , where
1 1
Sh = Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ) (12.42)
2 2
12.4 Robin Boundary Conditions 287
is the discrete Steklov–Poincaré operator which is defined by using the

matrices
Dh [j, i] = Dϕ1i , ϕ1j ΓN , Kh [, i] = Kϕ1i , ϕ0 Γ ,
Vh [, k] = V ϕ0k , ϕ0 Γ , Mh [, i] = ϕ1i , ϕ0 Γ
for i, j = 1, . . . , MN and k, = 1, . . . , N . By using
1
w = Vh−1 ( Mh + Kh )û
2
the linear system Sh û = f is equivalent to

& '& ' & '
Vh − 12 Mh − Kh w 0
1
= . (12.43)
2 Mh + Kh Dh û f
For a comparison of the symmetric boundary integral formulation (12.31)

with the symmetric boundary element approximation (12.38) of the Steklov–
Poincaré operator, and to check the theoretical error estimates we now con-
sider a simple model problem. Let the domain Ω = (0, 0.5)2 be a square with
the boundary Γ = ∂Ω which is decomposed into a Dirichlet part
ΓD = Γ ∩ {x ∈ R2 : x2 = 0},
and the remaining Neumann part ΓN = Γ \ΓD . The boundary data gD and
gN are given in such a way such that the solution of the mixed boundary value
problem (7.35) is
1
u(x) = − log |x − x∗ |, x∗ = (−0.1, −0.1) .
2
The boundary Γ = ∂Ω is decomposed uniformly into N boundary elements
τ of the same mesh width h. The boundary conditions on ΓD and ΓN are
interpolated by using piecewise linear and piecewise constant basis functions,
respectively. In Table 12.4 we give the errors and the related order of con-
vergence for the symmetric formulation (12.31) which confirm the theoretical
error estimates (12.33) and (12.34). The same holds true for the error esti-
mates (12.40) and (12.41) in the case of the symmetric approximation, for
which the numerical results are given in Table 12.5.
12.4 Robin Boundary Conditions

Finally we consider the variational problem (7.41) of the homogeneous Robin
boundary value problem (1.10) and (1.13) to find γ0int u ∈ H 1/2 (Γ ) such that
Sγ0int u, vΓ + κγ0int u, vΓ = g, vΓ

N t − th L2 (Γ ) eoc u − uh L2 (Γ ) eoc

8 9.151 –1 7.623 –2
16 5.304 –1 0.79 2.305 –2 1.73
32 2.767 –1 0.94 6.166 –3 1.90
64 1.389 –1 0.99 1.544 –3 2.00
128 6.913 –2 1.01 3.776 –4 2.03
256 3.438 –2 1.01 9.180 –5 2.04
512 1.713 –2 1.01 2.234 –5 2.04
1024 8.544 –3 1.00 5.451 –6 2.04
2048 4.266 –3 1.00 1.328 –6 2.04
theory: 1 2
Table 12.4. Error and order of convergence, symmetric formulation.
N t − th L2 (Γ ) eoc u − uh L2 (Γ ) eoc

8 8.725 –1 6.437 –2
16 5.140 –1 0.76 1.832 –2 1.81
32 2.710 –1 0.92 4.851 –3 1.92
64 1.370 –1 0.98 1.244 –3 1.96
128 6.855 –2 1.00 3.143 –4 1.98
256 3.422 –2 1.00 7.893 –5 1.99
512 1.708 –2 1.00 1.977 –5 2.00
1024 8.531 –3 1.00 4.948 –6 2.00
2048 4.263 –3 1.00 1.239 –6 2.00
theory: 1 2
Table 12.5. Error and order of convergence, symmetric approximation.
is satisfied for all v ∈ H 1/2 (Γ ). By using the ansatz

M
uh (x) = ui ϕ1i (x) ∈ Sh1 (Γ )
i=1
the corresponding Galerkin variational formulation is equivalent to a linear

system of algebraic equations,
0 κ 1
M̄h + Sh u = f , (12.44)
where
Sh [j, i] = Sϕ1i , ϕ1j Γ , M̄hκ [j, i] = κϕ1i , ϕ1j Γ , fj = g, ϕ1j Γ
and i, j = 1, . . . , M . As in the case when considering a mixed boundary value

problem we need to introduce a suitable approximation of the Galerkin stiff-
ness matrix Sh of the Steklov Poincaré operator S. Then, instead of (12.44)
we have to solve
12.5 Exercises 289

κ 1 −1 1
M̄h + Dh + ( Mh + Kh )Vh ( Mh + Kh ) û = f ,
2 2
or the equivalent system

& '& ' & '
Vh − 12 Mh − Kh w 0
1 κ
= .
2 Mh + Kh M̄h + Dh û f
12.5 Exercises
12.1 Let τk be a one–dimensional boundary element of length hk with mid-
point x∗k . Compute the collocation matrix element

1
VhC [k, k] = − log |y − x∗k |dsy .
2π
τk
12.2 Let τk be a one–dimensional boundary element of length hk . Compute

the Galerkin matrix element

1
VhG [k, k] = − log |x − y|dsy dsx .
2π
τk τk
12.3 Using a simple midpoint rule the Galerkin matrix element of Exercise
12.2 may be approximated by
VhG [k, k] ≈ hk VhC [k, k].
Compute the relative error
|VhG [k, k] − hk VhC [k, k]|

VhG [k, k]
and investigate the behavior as hk → 0.

12.4 Give an explicite integration formula for the pointwise evaluation of the
double layer potential of the Laplace operator when using piecewise linear
basis functions.
12.5 Let τk be some plane boundary element in R3 . For a function u(x) defined
for x ∈ τk the extension u is defined to be constant along the normal vector
n(x). The surface curl is then given as
curlτk u(x) = n(x) × ∇x u

(x).
Compute the surface curl when u(x) is linear on τk .

12.6 Let Γ = ∂Ω be the boundary of the circle Ω = Br (0) ⊂ R2 which can

be described by using polar coordinates. Describe the Galerkin discretization
of the Laplace single layer potential (cf. Exercise 6.1) when using piecewise
constant basis functions with respect to a uniform decomposition of the pa-
rameter domain [0, 1]. In particular prove that Vh is a circulant matrix, i.e.
Vh [ + 1, k + 1] = Vh [, k] for k, = 1, . . . , N − 1
and
Vh [ + 1, 1] = Vh [, N ] for = 1, . . . , N − 1.
12.7 Compute all eigenvalues and eigenvectors of the matrix

⎛ ⎞
0 1 0 ··· ··· 0
⎜0 0 1 0 ··· 0⎟
⎜ ⎟
⎜ .. . . . . . . .. ⎟
⎜. . . . ⎟
J = ⎜ . ⎟ ∈ RN ×N
⎜0 ··· ··· 0 1 0⎟
⎜ ⎟
⎝0 ··· ··· ··· 0 1⎠
1 0 ··· ··· ··· 0
as well as of J , = 2, . . . , N .
12.8 Compute all eigenvalues and eigenvectors of a circulant matrix A. De-
scribe the inverse matrix A−1 .
13
Iterative Solution Methods
The Galerkin discretization of variational problems as described in Chapter 8

leads to large linear systems of algebraic equations. In the case of an elliptic
and self adjoint partial differential operator the system matrix is symmetric
and positive definite. Therefore we may use the method of conjugate gradients
to solve the resulting system iteratively. Instead, the Galerkin discretization of
a saddle point problem, e.g. when considering a mixed finite element scheme
or the symmetric formulation of boundary integral equations, leads to a linear
system where the system matrix is positive definite but block skew symmetric.
By applying an appropriate transformation this system can be solved again
by using a conjugate gradient method. Since we are interested in iterative
solution algorithms where the convergence behavior is independent of the
problem size, i.e. which is robust with respect to the mesh size, we need to
use appropriate preconditioning strategies. For this we describe and analyze
first a quite general approach which is based on the use of operators of the
opposite order, and give later two examples for both finite and boundary
element methods. For a more detailed theory of general iterative methods we
refer to [4, 11, 70, 143].
13.1 The Method of Conjugate Gradients
We need to compute the solution vectors u ∈ RM of a sequence of linear

systems of algebraic equations (8.5), AM u = f , where the system matrix
AM ∈ RM ×M is symmetric and positive definite, and where M ∈ N is the
dimension of the trial space to be used for the discretization of the underlying
elliptic variational problem (8.1).
To derive the method of conjugate gradients we start with a system of
−1
conjugate or AM –orthogonal vectors {pk }M k=0 satisfying
(AM pk , p ) = 0 for k, = 0, . . . , M − 1, k = .
292 13 Iterative Solution Methods
Since the system matrix AM is supposed to be positive definite, we have
(AM pk , pk ) > 0 for k = 0, 1, . . . , M − 1.
For an arbitrary given initial guess u0 ∈ RM we can write the unique solution
u ∈ RM of the linear system AM u = f as a linear combination of conjugate
vectors as
M−1
u = u0 − α p .
=0
Hence we have

M −1
AM u = AM u0 − α AM p = f ,
=0
and from the AM –orthogonality of the basis vectors p we can compute the
yet unknown coefficients from
(AM u0 − f , p )
α = for = 0, 1, . . . , M − 1.
(AM p , p )
For some k = 0, 1, . . . , M we may define an approximate solution

k−1
uk := u0 − α p ∈ RM
=0
of the linear system AM u = f . Obviously, uM = u is just the exact solution.

By construction we have
uk+1 := uk − αk pk for k = 0, 1, . . . , M − 1,
−1
and from the AM –orthogonality of the vectors {p }M
=0 we obtain
& '

k−1
0
AM u − α AM p − f , p
k
(AM u0 − f , pk ) =0 (AM uk − f , pk )
αk = = = .
(AM pk , pk ) (AM pk , pk ) (AM pk , pk )
If we denote by
rk := AM uk − f
the residual of the approximate solution uk we finally have
(rk , pk )
αk = . (13.1)
(AM pk , pk )
On the other hand, for k = 0, 1, . . . , M − 1 we can compute the residual rk+1

recursively by
13.1 The Method of Conjugate Gradients 293
rk+1 = AM uk+1 − f = AM (uk − αk pk ) − f = rk − αk AM pk .
The above approach is based on the use of AM –orthogonal vectors {p }M 1

=0 .
Such a vector system can be constructed by applying the Gram–Schmidt
−1
orthogonalization algorithm which is applied to some given system {w }M
=0
of vectors which are linear independent, see Algorithm 13.1.
Initialize for k = 0:
p0 := w0
Compute for k = 0, 1, . . . , M − 2:
k
(AM wk+1 , p )
pk+1 := wk+1 − βk p , βk =
(Ap , p )
=0
Algorithm 13.1: Gram–Schmidt orthogonalization.
By construction we have for k = 0, 1, . . . , M − 1
span {p }k =0 = span {w }k =0 .
−1
It remains to define the initial vector system {w }M
=0 . One possibility is to
k k M
choose the unit basis vectors w := e = (δk+1, ) =1 [59]. Alternatively we
may find the basis vector wk+1 from the properties of the already constructed
vector systems {p }k =0 and {r }k =0 .
Lemma 13.1. For k = 0, 1, . . . , M − 2 we have
(rk+1 , p ) = 0 for = 0, 1, . . . , k.
Proof. For = k = 1, . . . , M − 1 we have by using (13.1) to define the coeffi-

cients αk and by using the recursion of the residual rk+1 the orthogonality
(rk+1 , pk ) = (rk , pk ) − αk (AM pk , pk ) = 0 .
For = k − 1 we then obtain
(rk+1 , pk−1 ) = (rk , pk−1 ) − αk (AM pk , pk−1 ) = 0
by applying the AM –orthogonality of pk and pk−1 . Now the assertion follows

by induction.
From the orthogonality relation between the residual rk+1 and the search
directions p we can immediately conclude a orthogonality of the residual rk+1
with the initial vectors w .
Corollary 13.2. For k = 0, 1, . . . , M − 2 we have
(rk+1 , w ) = 0 for = 0, 1, . . . , k.
Proof. By construction of the search directions p from the Gram–Schmidt

orthogonalization we first have the representation

−1
w = p + β −1,j pj .
j=0
Hence we obtain

−1
(rk+1 , w ) = (rk+1 , p ) + β −1,j (rk+1 , pj ),
j=0
and the orthogonality relation follows from Lemma 13.1.

Hence we have that all vectors
w0 , w1 , . . . , wk , rk+1
are orthogonal to each other, and therefore linear independent. Since we need
only to know the search directions p0 , . . . , pk and therefore the initial vec-
tors w0 , . . . , wk to construct the approximate solution uk+1 and therefore the
residual rk+1 , we can define the new initial vector as
wk+1 := rk+1 for k = 0, . . . , M − 2
where w0 := r0 . By Corollary 13.2 we then have the orthogonality
(rk+1 , r ) = 0 for = 0, . . . , k; k = 0, . . . , M − 2.
Moreover, for the numerator of the coefficient αk we obtain

k−1
(rk , pk ) = (rk , rk ) + βk−1, (rk , p ) = (rk , rk ),
=0
and therefore, instead of (13.1),
(rk , rk )
αk = for k = 0, . . . , M − 1.
(AM pk , pk )
In what follows we can assume
α > 0 for = 0, . . . , k.
Otherwise we would have
(r +1 , r +1 ) = (r − α Ap , r +1 ) = (r , r +1 ) = 0
implying r +1 = 0 and therefore u +1 = u would be the exact solution of the

linear system AM u = f .
From the recursion of the residual r +1 we then obtain

1
AM p = r − r +1 for = 0, . . . , k.
α
Now, by using wk+1 = rk+1 and by using the symmetry of the system matrix
AM = AM we can compute the nominator of the coefficient βk as
1 k+1
(AM wk+1 , p ) = (rk+1 , AM p ) = (r , r − r +1 )
α
⎧
⎪
⎨ 0 for < k,
= k+1
, rk+1 )
⎪
⎩−
(r
for = k.
αk
The recursion of the Gram–Schmidt orthogonalization algorithm now reduces
to
1 (rk+1 , rk+1 )
pk+1 = rk+1 − βkk pk where βkk = − .
αk (AM pk , pk )
On the other hand we have
αk (AM pk , pk ) = (rk −rk+1 , pk ) = (rk , pk ) = (rk , rk −βk−1,k−1 pk−1 ) = (rk , rk )
and therefore
(rk+1 , rk+1 )
pk+1 = rk+1 + βk pk where βk = .
(rk , rk )
Summarizing the above we obtain the iterative method of conjugate gradients

[78] as described in Algorithm 13.2.
For an arbitrary initial guess u0 compute

r0 := AM u0 − f , p0 := r0 ,
0 := (r0 , r0 ).
For k = 0, 1, 2, . . . , M − 2:
sk := AM pk , σk := (sk , pk ), αk :=
k /σk ;
uk+1 := uk − αk pk , rk+1 := rk − αk sk ;

k+1 := (rk+1 , rk+1 ) .
Stop, if
k+1 ≤ ε
0 is satisfied for some given accuracy ε.
Otherwise, compute the new search direction
βk :=
k+1 /
k , pk+1 := rk+1 + βk pk .
Algorithm 13.2: Method of conjugate gradients.

If the matrix AM is symmetric and positive definite we may define

.
· AM := (AM ·, ·)
to be an equivalent norm in RM . Moreover,
λmax (AM )
κ2 (AM ) := AM 2 A−1
M 2 =
λmin (AM )
is the spectral condition number of the positive definite and symmetric ma-
trix AM where · 2 is the matrix norm which is induced by the Euclidean
inner product. Then one can prove the following estimate for the approximate
solution uk , see for example [70, 143].
Theorem 13.3. Let AM = A M > 0 be symmetric and positive definite, and
let u ∈ RM be the unique solution of the linear system AM u = f . Then the
method of conjugate gradients as described in Algorithm 13.2 is convergent for
any initial guess u0 ∈ RM , and there holds the error estimate
.
2q k 0 κ2 (AM ) + 1
uk − uA ≤ u − uA where q := . .
1 + q 2k κ2 (AM ) − 1
To ensure a certain given relative accuracy ε ∈ (0, 1) we find the number
kε ∈ N of required iteration steps from
uk − uA 2q k
0
≤ ≤ ε
u − uA 1 + q 2k
and therefore √
ln[1 − 1 − ε2 ] − ln ε
kε > .
ln q
The number kε obviously depends on q and therefore on the spectral condi-
tion number κ2 (AM ) of AM . When considering the discretization of elliptic
variational problems by using either finite or boundary elements the spectral
condition number κ2 (AM ) depends on the dimension M ∈ N of the used finite
dimensional trial space, or on the underlying mesh size h.
In the case of a finite element discretization we have for the spectral con-
dition number, by using the estimate (11.13),
κ2 (AFEM
h ) = O(h−2 ), h/2 ) ≈ 4 κ2 (Ah
i.e. κ2 (AFEM FEM
)
when considering a globally quasi–uniform mesh refinement strategy. Asymp-
totically, this gives
/
κ2 (AFEM .
h/2 ) + 1 2 κ2 (AFEM )+1
ln qh/2 = ln / ≈ ln . h
κ2 (Ah/2 ) − 1
FEM 2 κ2 (Ah ) − 1
FEM
.
1 κ2 (AFEM )+1 1
≈ ln . h
= ln qh .
2 κ2 (Ah ) − 1
FEM 2
Therefore, in the case of an uniform refinement step, i.e. halving the mesh
size h, the number of required iterations is doubled to reach the same relative
accuracy ε. As an example we choose ε = 10−10 . In Table 13.1 we give the
number of iterations of the conjugate gradient method to obtain the results
which were already presented in Table 11.2.
FEM BEM
L N κ2 (Ah ) Iter N κ2 (Vh ) Iter
2 64 12.66 13 16 24.14 8
3 256 51.55 38 32 47.86 18
4 1024 207.17 79 64 95.64 28
5 4096 829.69 157 128 191.01 39
6 16384 3319.76 309 256 381.32 52
7 65536 13280.04 607 512 760.73 69
8 262144 52592.92 1191 1024 1516.02 91
Theory: O(h−2 ) O(h−1 ) O(h−1 ) O(h−1/2 )
Table 13.1. Number of CG iteration steps when ε = 10−10 .
When considering a comparable discretization by using boundary elements as

already discussed in Table 12.2 we obtain for the spectral condition number
of the system matrix
κ2 (ABEM
h ) = O(h−1 ) h/2 ) ≈ 2 κ2 (Ah
i.e. κ2 (ABEM BEM
).
The number of required√iterations to reach a certain relative accuracy ε then

grows with a factor of 2, see Table 13.1.
Hence there is a serious need to construct iterative algorithms which are
almost robust with respect to all discretization parameters, i.e. with respect
to the mesh size h. In general this can be done by introducing the concept of
preconditioning the linear system AM u = f .
Let CA ∈ RM ×M be a symmetric and positive definite matrix which can
be factorized as
CA = JDCA J , DCA = diag(λk (CA )), λk (CA ) > 0
where J ∈ RM ×M contains all eigenvectors of CA which are assumed to be

orthonormal. Hence we can define
1/2 1/2
.
CA = JDCA J , DCA = diag( λk (CA ))
satisfying
1/2 1/2 −1/2 1/2
CA = CA CA , CA := (CA )−1 .
Instead of the linear system AM u = f we now consider the equivalent system
u := C −1/2 AM C −1/2 C 1/2 u = C −1/2 f =: f

A A A A A
where the transformed system matrix
:= C −1/2 AM C −1/2
A A A
is again symmetric and positive definite. Hence we can apply the method
of conjugate gradients as described in Algorithm 13.2 to compute the trans-
1/2
= CA u. Inserting all the transformations we finally
formed solution vector u
obtain the preconditioned method of conjugate gradients, see Algorithm 13.3.
For an arbitrary initial guess x0 compute

−1 0
r0 := AM x0 − f , v 0 := CA r , p0 := v 0 ,
0 := (v 0 , r0 ).
For k = 0, 1, 2, . . . , M − 2:
sk := AM pk , σk := (sk , pk ), αk :=
k /σk ;
xk+1 := xk − αk pk , rk+1 := rk − αk sk ;
−1 k+1
v k+1 := CA r ,
k+1 := (v k+1 , rk+1 ) .
Stop, if
k+1 ≤ ε
Otherwise compute the new search direction
βk :=
k+1 /
k , pk+1 := v k+1 + βk pk .
Algorithm 13.3: Preconditioned method of conjugate gradients.
The Algorithm 13.3 of the preconditioned method of conjugate gradients re-

quires one matrix by vector product per iteration step, sk = AM pk , and one
−1 k+1
application of the inverse preconditioning matrix, v k+1 = CA r . From
Theorem 13.3 we obtain an error estimate for the approximate solution u k ,
/
2
q +1
κ2 (A)
0
uk − u
A ≤
u −
u where
q = / .
1 + q2k A
−1
κ2 (A)
1/2
Note that for z = CA z we have

z 2A = (AC

1/2 1/2 2
A z, CA z) = (AM z, z) = zAM .
−1/2 k
Hence, for the approximate solution uk = CA we find the error estimate
u
2
q
uk − uAM ≤ u0 − uAM .
1 + q2k
13.2 A General Preconditioning Strategy 299
we
To bound the extremal eigenvalues of the transformed system matrix A
get from the Rayleigh quotient
z , z) z , z)
= min (A
λmin (A) ≤ max
(A
= λmax (A).
z , z)
z∈RM ( z , z)
z∈RM (
= C −1/2 AM C −1/2 and z = C 1/2 z this

When inserting the transformations A A A A
gives
= min (AM z, z) ≤ max (AM z, z) = λmax (A).

λmin (A)
z∈RM (CA z, z) z∈RM (CA z, z)
Hence we have to assume that the preconditioning matrix CA satisfies the

spectral equivalence inequalities
1 (CA z, z) ≤ (AM z, z) ≤ c2 (CA z, z)

cA for z ∈ RM
A
(13.2)
independent of M . Then we can bound the spectral condition number of the

transformed system matrix as
−1/2 −1/2 −1 cA
2
κ2 (CA ACA ) = κ2 (CA AM ) ≤ .
cA
1
−1
If the spectral condition number κ2 (CA AM ) of the preconditioned system
matrix can be bounded independent of the dimension M , i.e. independent of
the mesh size h, then there is a fixed number kε of required iterations to reach
a certain given relative accuracy ε.
13.2 A General Preconditioning Strategy

We need to construct a matrix CA as a preconditioner for a given matrix
AM such that the spectral equivalence inequalities (13.2) are satisfied, and an
−1 k
efficient realization of the preconditioning v k = CA r is possible. Here we
consider the case where the matrix AM represents a Galerkin discretization
of a bounded, X–elliptic, and self–adjoint operator A : X → X satisfying
2
Av, v ≥ cA
1 vX , AvX ≤ cA
2 vX for v ∈ X. (13.3)
In particular, the matrix AM is given by
AM [, k] = Aϕk , ϕ for k, = 1, . . . , M
where XM := span{ϕk }M k=1 ⊂ X is some conforming trial space.

Let B : X → X be some bounded, X –elliptic, and self–adjoint operator,
i.e. for f ∈ X we assume
2
Bf, f ≥ cB
1 f X , Bf X ≤ cB
2 f X .
By applying Theorem 3.4 there exists the inverse operator B −1 : X → X . In

particular, by using (3.13) and Lemma 3.5 we have
1 1
B −1 vX ≤ vX , B −1 v, v ≥ v2X for v ∈ X. (13.4)
cB
1 cB
2
From the assumptions (13.3) and (13.4) we immediately conclude:

Corollary 13.4. For v ∈ X there hold the spectral equivalence inequalities
−1 −1
1 c1 B
cA v, v ≤ Av, v ≤ cA
2 c2 B
B B
v, v.
Then, by defining the preconditioning matrix
CA [, k] = B −1 ϕk , ϕ for k, = 1, . . . , M (13.5)
we obtain from Corollary 13.4 by using the isomorphism

M
v∈R M
↔ vM = v k ϕk ∈ X M ⊂ X
k=1
the required spectral equivalence inequalities
1 c1 (CA v, v) ≤ (AM v, v) ≤ c2 c2 (CA v, v)

cA for v ∈ RM .
B A B
(13.6)
Although the constants in (13.6) only express the continuous mapping prop-
erties of the operators A and B, and therefore they are independent of the
discretization to be used, the above approach seems on a first glance useless,
since in general only the operator B is given explicitly. Moreover, neither can
−1
the preconditioning matrix CA be computed nor can the inverse CA be ap-
plied efficiently. Hence we introduce a conforming trial space in the dual space
X ,

XM := span{ψk }Mk=1 ⊂ X ,
and define
BM [, k] = Bψk , ψ , MM [, k] = ϕk , ψ for k, = 1, . . . , M.
Note that the Galerkin matrix BM is symmetric and positive definite, and
therefore invertible. Therefore we can define an approximation of the precon-
ditioning matrix CA by
A := M B −1 MM .
C (13.7)
M M
A is spec-
We need to prove that the approximated preconditioning matrix C
trally equivalent to CA , and therefore to AM .
Lemma 13.5. Let CA be the Galerkin matrix of B −1 as defined in (13.5),
A be the approximation as given in (13.7). Then there holds
and let C
A v, v) ≤ (CA v, v)
(C for v ∈ RM .
Proof. Let v ∈ RM ↔ vM ∈ XM ⊂ X be arbitrary but fixed. Then,

w = B −1 vM ∈ X is the unique solution of the variational problem
Bw, z = vM , z for z ∈ X .
Note that
(CA v, v) = B −1 vM , vM = w, vM = Bw, w. (13.8)
−1
In the same way we define w = BM MM v ↔ wM ∈ XM as the unique solution
of the Galerkin variational problem

BwM , zM = vM , zM for zM ∈ XM ,
Again,
A v, v) = (B −1 MM v, MM v) = (w, MM v) = wM , vM = BwM , wM .
(C M
(13.9)
Moreover we have the Galerkin orthogonality

B(w − wM ), zM = 0 for zM ∈ XM .
By using the X –ellipticity of B we now have

2
0 ≤ cB
1 w − wM X ≤ B(w − wM ), w − wM
= B(w − wM ), w = Bw, w − BwM , wM
and therefore
BwM , wM ≤ Bw, w.
By using (13.8) and (13.9) this finally gives the assertion.
Note that Lemma 13.5 holds for any arbitrary conforming trial spaces

XM ⊂ X and XM ⊂ X . However, to prove the reverse estimate we need to

assume a certain stability condition of the trial space XM ⊂ X .
Lemma 13.6. In addition to the assumptions of Lemma 13.5 we assume the

stability condition
vM , zM
cS vM X ≤ sup for all vM ∈ XM . (13.10)

0=zM ∈XM zM X
Then,
2
cB
1 A v, v)
cS (CA v, v) ≤ (C for all v ∈ RM .
cB
2
Proof. Let v ∈ RM ↔ vM ∈ XM be arbitrary but fixed. From the properties

(13.4) we then obtain
1
(CA v, v) = B −1 vM , vM ≤ B −1 vM X vM X ≤ vM 2X .
cB
1
−1
As in the proof of Lemma 13.5 let w = BM MM v ↔ wM ∈ XM . Then, by
using the stability assumption (13.10),
vM , zM BwM , zM
cS vM X ≤ sup = sup ≤ cB
2 wM X ,

0=zM ∈XM zM X
0=zM ∈XM zM X
and hence
2 2
1 cB
2 1 cB
2
(CA v, v) ≤ wM 2X ≤ BwM , wM .
cB
1 cS cS cB
1
By using (13.9) this gives the assertion.

Together with (13.6) we now conclude the spectral equivalence inequalities
A and AM .
of C
Corollary 13.7. Let all assumptions of Lemma 13.6 be satisfied, in partic-
ular we assume the stability condition (13.10). Then there hold the spectral
equivalence inequalities
2
A B 1 cB
2 A v, v) for all v ∈ RM .
c1 c1 (CA v, v) ≤ (AM v, v) ≤ c2 c2
A B
(C
cS cB
1

Due to dim XM = dim XM the discrete stability condition (13.10) also ensures
the invertibility of the matrix MM . Hence for the inverse of the approximated
preconditioning matrix C A we obtain
−1 = M −1 BM M − ,
CA M M

in particular we need to invert sparse matrices MM and MM , and in addition
we have to perform one matrix by vector multiplication with BM .
13.2.1 An Application in Boundary Element Methods
The general approach of preconditioning as described in Section 13.2 is now

applied to construct some preconditioners to be used in boundary element
methods. By considering the single layer potential V : H −1/2 (Γ ) → H 1/2 (Γ )
and the hypersingular boundary integral operator D : H 1/2 (Γ ) → H −1/2 (Γ )
there is given a suitable pair of boundary integral operators of opposite order
[104, 105, 144]. However, the hypersingular boundary integral operator D is
±1/2
only semi–elliptic, hence we have to use appropriate factor spaces H∗ (Γ )
as already considered in Section 6.6.1.
Lemma 13.8. For the single layer potential V and for the hypersingular
boundary integral operator D there hold the spectral equivalence inequalities
−1 1 −1
1 V
cV1 cD v, vΓ ≤ D
v , vΓ ≤ V v, vΓ
4
1/2
for all v ∈ H∗ (Γ ) = {v ∈ H 1/2 (Γ ) : v, weq Γ = 0}.
−1/2 1/2
Proof. The single layer potential V : H∗ (Γ ) → H∗ (Γ ) defines an iso-
1/2
morphism. Hence, for an arbitrary given v ∈ H∗ (Γ ) there exists a unique
−1/2
∈ H∗
w (Γ ) such that v = V w.
By using the symmetry relations (6.25)
and (6.26) of all boundary integral operators we obtain the upper estimate,
v , vΓ = DV w,
D V w
Γ
1 1
= ( I − K )( I + K )w,
V w
Γ
2 2
1 1
= ( I + K )w,
( I − K)V w Γ
2 2
1 1
= ( I + K )w,
V ( I − K )w
Γ
2 2
1
= V w, Γ − V K w,
w K w
Γ
4
1 1
≤ V w, Γ = V −1 v, vΓ .
w
4 4
From the H −1/2 (Γ )–ellipticity of the single layer potential V (see Theorem
6.22 in the three–dimensional case d = 3, and Theorem 6.23 in the two–
dimensional case d = 2) we conclude, by using the estimate (3.13), the bound-
edness of the inverse single layer potential,
1
V −1 v, vΓ ≤ v2H 1/2 (Γ ) for all v ∈ H 1/2 (Γ ).
cV1
1/2
By using the H∗ (Γ )–ellipticity of the hypersingular boundary integral op-
erator D (see Theorem 6.24) we then obtain the lower estimate
v , vΓ ≥
D v 2H 1/2 (Γ ) ≥ cD
1
cD 1 c1 V
V −1
v, vΓ
1/2
for all v ∈ H∗ (Γ ).
By the bilinear form
vΓ := Du, vΓ + α u, weq Γ v, weq Γ
Du,
for u, v ∈ H 1/2 (Γ ) we may define the modified hypersingular boundary inte-
gral operator D : H 1/2 (Γ ) → H −1/2 (Γ ) where α ∈ R+ is some parameter to
be chosen appropriately, and weq = V −1 1 ∈ H −1/2 (Γ ) is the natural density.
Theorem 13.9. For the single layer potential V and for the modified hy-
there hold the spectral equivalence
persingular boundary integral operator D
inequalities
vΓ ≤ γ2 V −1 v, vΓ
γ1 V −1 v, vΓ ≤ Dv, (13.11)
for all v ∈ H 1/2 (Γ ) where

1
1 , α1, weq Γ ,
γ1 := min cV1 cD γ2 := max , α1, weq Γ .
4
Proof. For any v ∈ H 1/2 (Γ ) we consider the orthogonal decomposition
v, weq Γ 1/2

v = v + γ, γ := , v ∈ H∗ (Γ ).
1, weq Γ
The bilinear form of the inverse single layer potential can then be written as
[v, weq Γ ]2
V −1 v, vΓ = V −1 v, vΓ + .
1, weq Γ
By using Lemma 13.8 we now obtain

vΓ = D
Dv, v , vΓ + α [v, weq Γ ]2
1 −1 [v, weq Γ ]2
≤ V v, vΓ + α 1, weq Γ
4 1, weq Γ

1
≤ max , α1, weq Γ V −1 v, vΓ .
4
The lower estimate follows in the same way.

From the previous theorem we can find an optimal choice of the positive
parameter α ∈ R+ .
Corollary 13.10. When choosing

1
α :=
41, weq Γ
we obtain the spectral equivalence inequalities
−1 vΓ ≤ 1 −1
1 V
cV1 cD v, vΓ ≤ Dv, V v, vΓ
4
for all v ∈ H 1/2 (Γ ).
By using Corollary 13.10 we now can define a preconditioner for the linear
system (12.15) of the Dirichlet boundary value problem, and for the sys-
tem (12.27) of the Neumann boundary value problem. The system matrix in
h := Dh + α a a where
(12.27) is D
Dh [j, i] = Dϕ1i , ϕ1j Γ , aj = ϕ1j , weq Γ
for i, j = 1, . . . , M and ϕ1i ∈ Sh1 (Γ ) are piecewise linear continuous basis

functions. In addition we define
V̄h [j, i] = V ϕ1i , ϕ1j Γ , M̄h [j, i] = ϕ1i , ϕ1j Γ
for i, j = 1, . . . , M .
Lemma 13.11. Let the L2 projection Qh : H 1/2 (Γ ) → Sh1 (Γ ) ⊂ H 1/2 (Γ ) be

bounded. Then there holds the stability condition
1 vh , wh Γ
vh H 1/2 (Γ ) ≤ sup for all vh ∈ Sh1 (Γ ).
cQ 1 (Γ ) wh H −1/2 (Γ )
0=wh ∈Sh
Proof. The L2 projection Qh : H 1/2 (Γ ) → Sh1 (Γ ) ⊂ H 1/2 (Γ ) is bounded, i.e.
Qh vH 1/2 (Γ ) ≤ cQ vH 1/2 (Γ ) for all v ∈ H 1/2 (Γ ).
For any w ∈ H −1/2 (Γ ) the L2 projection Qh w ∈ Sh1 (Γ ) is defined as the

unique solution of the variational problem
Qh w, vh L2 (Γ ) = w, vh Γ for all vh ∈ Sh1 (Γ ).
Then,
Qh w, vΓ Qh w, Qh vL2 (Γ )

Qh wH −1/2 (Γ ) = sup = sup
0=v∈H 1/2 (Γ ) vH 1/2 (Γ ) 0=v∈H 1/2 (Γ ) vH 1/2 (Γ )
w, Qh vΓ Qh vH 1/2 (Γ )
= sup ≤ wH −1/2 (Γ ) sup
0=v∈H 1/2 (Γ ) vH 1/2 (Γ ) 0=v∈H 1/2 (Γ ) vH 1/2 (Γ )
≤ cQ wH −1/2 (Γ ) ,
which implies the boundedness of Qh : H −1/2 (Γ ) → Sh1 (Γ ) ⊂ H −1/2 (Γ ). Now

the stability estimate follows by applying Lemma 8.5.
By using Lemma 13.11 all assumptions of Lemma 13.6 are satisfied, i.e.
CD := M̄h V̄h−1 M̄h
h . In par-
defines a preconditioning matrix which is spectrally equivalent to D
ticular there hold the spectral equivalence inequalities
2
h v, v) ≤ 1 cV2
cV1 cD
1 (CD v, v) ≤ (D cQ V (CD v, v) for all v ∈ RM .
4 c1
In Table 13.2 the extremal eigenvalues and the resulting spectral condition
−1
numbers of the preconditioned system matrix CD Dh are listed for the L–
shaped domain as given in Fig. 10.1. For comparison we also give the corre-
sponding values in the case of a simple diagonal preconditioning which show
a linear dependency on the inverse mesh parameter h−1 .
By applying Corollary 13.10 we can use the Galerkin discretization of
the modified hypersingular boundary integral operator D as a preconditioner
for the discrete single layer potential Vh in (12.15). However, when using
piecewise constant basis functions to discretize the single layer potential, for
the Galerkin discretization of the hypersingular boundary integral operator
CD
- = diag Dh CD −1
- = M̄h V̄h M̄h
−1 −1
L N λmin λmax κ(CD - Dh ) λmin λmax κ(CD - Dh )
0 28 9.05 –3 2.88 –2 3.18 1.02 –1 2.56 –1 2.50
1 112 4.07 –3 2.82 –2 6.94 9.24 –2 2.66 –1 2.88
2 448 1.98 –3 2.87 –2 14.47 8.96 –2 2.82 –1 3.14
3 1792 9.84 –3 2.90 –2 29.52 8.86 –2 2.89 –1 3.26
4 7168 4.91 –3 2.91 –2 59.35 8.80 –2 2.92 –1 3.31
5 28672 2.46 –4 2.92 –2 118.72 8.79 –2 2.92 –1 3.32
6 114688 1.23 –4 2.92 –2 237.66 8.78 –2 2.92 –1 3.33
Theory: O(h−1 ) O(1)
Table 13.2. Extremal eigenvalues and spectral condition number (BEM).
D requires the use of globally continuous basis functions. Moreover, as an

assumption of Lemma 13.6 we need to guarantee a related stability condition,
too. One possibility is to use locally quadratic basis functions [144]. For the
analysis of boundary integral preconditioners in the case of open curves, see
[104].
13.2.2 A Multilevel Preconditioner in Finite Element Methods
For u, v ∈ H 1 (Ω) we consider the bilinear form

a(u, v) = γ0int u(x)dsx γ0int v(x)dsx + ∇u(x)∇v(x)dx
Γ Γ Ω
which induces a bounded and H 1 (Ω)–elliptic operator A : H 1 (Ω) → H −1 (Ω).

This bilinear form is either related to the stabilized variational formulation
(4.31) of the Neumann boundary value problem, or to the variational formu-
lation of the Robin boundary value problem, or to the modified saddle point
formulation (4.22) when using Lagrange multipliers.
Let us assume that there is given a sequence {TNj }j∈N0 of globally quasi–
uniform decompositions of a bounded domain Ω ⊂ Rd where the global mesh
size hj of a decomposition TNj satisfies
c1 2−j ≤ hj ≤ c2 2−j (13.12)
for all j = 0, 1, 2, . . . with some global constants c1 and c2 . In particular, this

condition is satisfied when applying a globally uniform refinement strategy to
a given uniform coarse decomposition TN0 .
For each decomposition TNj the associated trial space of piecewise linear
continuous basis functions is given by
Vj := Sh1j (Ω) = span{ϕjk }k=1 ⊂ H 1 (Ω),

M
j
j ∈ N0 .
By construction we have
V0 ⊂ V1 ⊂ · · · ⊂ VL = Xh = Sh1L (Ω) ⊂ VL+1 ⊂ · · · ⊂ H 1 (Ω)
where Xh = Sh1 (Ω) ⊂ H 1 (Ω) is the trial space to be used for the Galerkin
−1 (Ω) which is induced by the
discretization of the operator A : H 1 (Ω) → H
bilinear form a(·, ·), i.e.
AhL [, k] = a(ϕL L

k , ϕ ) for k, = 1, . . . , ML .
It remains to construct a preconditioning matrix C A which is spectrally

equivalent to Ah . For this we need to have a preconditioning operator
B : H −1 (Ω) → H 1 (Ω) which satisfies the spectral equivalence inequalities
2 2
1 f H −1 (Ω) ≤ Bf, f Ω ≤ c2 f H −1 (Ω)
cB B
(13.13)
for all f ∈ H −1 (Ω) with some positive constants cB and cB . Such an op-
1 2
erator can be constructed when using an appropriately weighted multilevel
representation of L2 projection operators, see [28, 162].
For any trial space Vj ⊂ H 1 (Ω) let Qj : L2 (Ω) → Vj be the L2 projection
operator as defined in (9.23), i.e. Qj u ∈ Vj is the unique solution of the
variational problem
Qj u, vj L2 (Ω) = u, vj L2 (Ω) for all vj ∈ Vj .
Note that there holds the error estimate (9.28),
(I − Qj )uL2 (Ω) ≤ c hj |u|H 1 (Ω) for all u ∈ H 1 (Ω). (13.14)
In addition we assume an inverse inequality (9.19) to hold uniformly for all

trial spaces Vj , i.e.,
vj H 1 (Ω) ≤ cI h−1
j vj L2 (Ω) for all vj ∈ Vj . (13.15)
Finally, for j = −1 we define Q−1 := 0.
Lemma 13.12. For the sequence {Qj }j∈N0 of L2 projection operators Qj we

have the following properties:
1. Qk Qj = Qmin{k,j} ,
2. (Qk − Qk−1 )(Qj − Qj−1 ) = 0 for k = j,
3. (Qj − Qj−1 )2 = Qj − Qj−1 .
Proof. For uj ∈ Vj we have Qj vj = vj ∈ Vj and therefore Qj Qj v = Qj v for

all v ∈ L2 (Ω). In the case j < k we find Vj ⊂ Vk . Then, Qj v ∈ Vj ⊂ Vk and
thus Qk Qj v = Qj v. Finally, for j > k we obtain
Qj v, vj L2 (Ω) = v, vj L2 (Ω) for all vj ∈ Vj
and therefore
Qk Qj v, vk L2 (Ω) = Qj v, vk L2 (Ω) = v, vk L2 (Ω) = Qk v, vk L2 (Ω)
is satisfied for all vk ∈ Vk ⊂ Vj . This concludes the proof of 1. To show 2. we

assume j < k and therefore j ≤ k − 1. Then, by using 1. we obtain
(Qk − Qk−1 )(Qj − Qj−1 ) = Qk Qj − Qk−1 Qj − Qk Qj−1 + Qk−1 Qj−1

= Qj − Qj − Qj−1 + Qj−1 = 0.
By using 1. we finally get
(Qj − Qj−1 )2 = Qj Qj − Qj Qj−1 − Qj−1 Qj + Qj−1 Qj−1

= Qj − Qj−1 − Qj−1 + Qj−1 = Qj − Qj−1 .

By considering a weighted linear combination of L2 projection operators Qk

we define the multilevel operator
∞

1
B := h−2
k (Qk − Qk−1 ) (13.16)
k=0
which induces an equivalent norm in the Sobolev space H 1 (Ω).
Theorem 13.13. For the multilevel operator B 1 as defined in (13.16) there

hold the spectral equivalence inequalities
2 1 2
1 vH 1 (Ω) ≤ B v, vL2 (Ω) ≤ c2 vH 1 (Ω)
cB B
for all v ∈ H 1 (Ω).
The proof of Theorem 13.13 is based on several results. First we consider a

consequence of Lemma 13.12:
Corollary 13.14. For v ∈ H 1 (Ω) we have the representation

∞

B 1 v, vL2 (Ω) = h−2 2
k (Qk − Qk−1 )vL2 (Ω) .
k=0
Proof. By the definition of B 1 and by using Lemma 13.12, 3., we have

∞

B 1 v, vL2 (Ω) = h−2
k (Qk − Qk−1 )v, vL2 (Ω)
k=0
∞
= h−2 2
k (Qk − Qk−1 ) v, vL2 (Ω)
k=0
∞
= h−2
k (Qk − Qk−1 )v, (Qk − Qk−1 )vL2 (Ω)
k=0
∞
= h−2 2
k (Qk − Qk−1 )vL2 (Ω) .

k=0
From the inverse inequalities of the trial spaces Vk and by using the error
estimates of the L2 projection operators Qk we further obtain from Corollary
13.14:
Lemma 13.15. For all v ∈ H 1 (Ω) there hold the spectral equivalence inequal-
ities
∞
∞

c1 (Qk −Qk−1 )v2H 1 (Ω) ≤ B 1 v, vL2 (Ω) ≤ c2 (Qk −Qk−1 )v2H 1 (Ω) .
k=0 k=0
Proof. By using Lemma 13.12, 3., the triangle inequality, the error estimate
(13.14), and assumption (13.12) we have
∞

B 1 v, vL2 (Ω) = h−2 2
k (Qk − Qk−1 )vL2 (Ω)
k=0
∞

= h−2 2
k (Qk − Qk−1 )(Qk − Qk−1 )vL2 (Ω)
k=0
∞ !
≤ 2 h−2
k (Qk − I)(Qk − Qk−1 )v2L2 (Ω)
k=0
"
+ (I − Qk−1 )(Qk − Qk−1 )v2L2 (Ω)
∞
! "
≤ 2c h−2
k h2k (Qk − Qk−1 )v2H 1 (Ω) + h2k−1 (Qk − Qk−1 )v2H 1 (Ω)
k=0
∞

≤ c2 (Qk − Qk−1 )v2H 1 (Ω)
k=0
and therefore the upper estimate. To prove the lower estimate we get from
the global inverse inequality (13.15) for (Qk − Qk−1 )v ∈ Vk−1 , and by using
assumption (13.12),
∞
∞

(Qk − Qk−1 )v2H 1 (Ω) ≤ c2I h−2 2
k−1 (Qk − Qk−1 )vL2 (Ω)
k=0 k=0
∞

≤c h−2 2
k (Qk − Qk−1 )vL2 (Ω)
k=0
1
= c B v, vL2 (Ω) .

The statement of Theorem 13.13 now follows from Lemma 13.15 and from the
following spectral equivalence inequalities.
Lemma 13.16. For all v ∈ H 1 (Ω) there hold the spectral equivalence inequal-
ities
∞

c̄1 v2H 1 (Ω) ≤ (Qk − Qk−1 )v2H 1 (Ω) ≤ c̄2 v2H 1 (Ω) .
k=0
To prove Lemma 13.16 we first need a tool to estimate some matrix norms.
Lemma 13.17 (Schur Lemma). For a countable index set I we consider

the matrix A = (A[, k])k, ∈I and the vector u = (uk )k∈I . For an arbitrary
α ∈ R we then have
# $# $

2
Au2 ≤ sup |A[, k]| 2α(k− )
sup |A[, k]| 2α( −k)
u22 .
∈I k∈I
k∈I ∈I
Proof. Let v = Au. For an arbitrary ∈ I we first have

|v | = A[, k]uk ≤ |A[, k]| · |uk |

k∈I k∈I
. .
= |A[, k]| 2α(k− )/2 |A[, k]| 2α( −k)/2 |uk |.
k∈I
By applying the Cauchy–Schwarz inequality this gives

# $# $

|v |2 ≤ |A[, k]| 2α(k− ) |A[, k]| 2α( −k) u2k .
k∈I k∈I
Hence we have
# $# $

|v |2 ≤ |A[, k]| 2α(k− ) |A[, k]| 2α( −k) u2k
∈I ∈I k∈I k∈I
# $ # $

≤ sup |A[, k]| 2
α(k− )
|A[, k]| 2
α( −k)
u2k
∈I
k∈I ∈I k∈I
# $ # $

= sup |A[, k]| 2
α(k− )
|A[, k]| 2
α( −k)
u2k
∈I
k∈I k∈I ∈I
# $ # $

≤ sup |A[, k]| 2α(k− ) sup |A[, k]| 2α( −k) u2k
∈I k∈I
k∈I ∈I k∈I
which concludes the proof.

As a consequence of Lemma 13.17 we immediately obtain the norm esti-
mate
# $1/2 # $1/2

A2 ≤ sup |A[, k]| 2α(k− )
sup |A[, k]| 2α( −k)
(13.17)
∈I k∈I
k∈I ∈I
where α ∈ R is arbitrary. In particular for a symmetric matrix A and when

considering α = 0 the estimate

A2 ≤ sup |A[, k]| (13.18)
∈I
k∈I
follows. To prove the lower estimate in the spectral equivalence inequalities of

Lemma 13.16 we need to have a strengthened Cauchy–Schwarz inequality.
Lemma 13.18 (Strengthened Cauchy–Schwarz Inequality). Let as-

sumption (13.12) be satisfied. Then there exists a q < 1 such that

(Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω)
≤ c q |i−j| (Qi − Qi−1 )vH 1 (Ω) (Qj − Qj−1 )vH 1 (Ω)
holds for all v ∈ H 1 (Ω).
Proof. Without loss of generality we may assume j < i. For vj ∈ Vj we have

for the H 1 projection Q1j vj = vj ∈ Vj and therefore
(Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω) = (Qi − Qi−1 )v, Q1j (Qj − Qj−1 )vH 1 (Ω)
= Q1j (Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω)
≤ Q1j (Qi − Qi−1 )vH 1 (Ω) (Qj − Qj−1 )vH 1 (Ω) .
Due to Vj = Sh1j (Ω) ⊂ H 1+σ (Ω) the H 1 projection as given in (9.29) is well
defined for u ∈ H 1−σ (Ω) and for σ ∈ (0, 12 ). Dependent on the regularity
of the computational domain Ω there exists an index s ∈ (0, σ] such that

Q1j : H 1−s (Ω) → Vj ⊂ H 1−s (Ω) is bounded, see Lemma 9.11. By using
the inverse inequality in Vj and by using the error estimate (9.37) of the L2
projection Qj we then have
Q1j (Qi − Qi−1 )vH 1 (Ω) ≤ cI h−s
j (Qi − Qi−1 )vH 1−s (Ω)
= cI h−s
j (Qi − Qi−1 )(Qi − Qi−1 )vH 1−s (Ω)
0
≤ cI h−s
j (Qi − I)(Qi − Qi−1 )vH 1−s (Ω)
1
+(I − Qi−1 )(Qi − Qi−1 )vH 1−s (Ω)
0 s 1
≤ c h−s
j hi + hsi−1 (Qi − Qi−1 )vH 1 (Ω)
≤
c 2s(j−i) (Qi − Qi−1 )vH 1 (Ω) .
With q := 2−s we obtain the strengthened Cauchy–Schwarz inequality.

Proof of Lemma 13.16: Let Q1j : H 1 (Ω) → Sh1j (Ω) ⊂ H 1 (Ω) be the H 1
projection as defined by the variational problem (9.29), in particular for a
given u ∈ H 1 (Ω) the projection Q1j u ∈ Vj is the unique solution of
Q1j u, vj H 1 (Ω) = u, vj H 1 (Ω) for all vj ∈ Vj .

Then, dependent on the regularity of the computational domain Ω, and by
applying Lemma 9.11, there exists an index s ∈ (0, 1], such that the following
error estimate holds,
(I − Q1h )uH 1−s (Ω) ≤ c hs uH 1 (Ω) .
As in Lemma 13.12 we also have
(Q1j − Q1j−1 )(Q1j − Q1j−1 ) = Q1j − Q1j−1 .
Therefore, for v ∈ H 1 (Ω) we obtain the representation
∞
∞

v = (Q1i − Q1i−1 )v = vi where vi := (Q1i − Q1i−1 )v.
i=0 i=0
For i < k we therefore have vi = (Q1i − Q1i−1 )v ∈ Vi−1 ⊂ Vk−1 , and thus
(Qk − Qk−1 )vi = 0. Hence, by interchanging the order of summation,
∞
∞
∞
(Qk − Qk−1 )v2H 1 (Ω) = (Qk − Qk−1 )vi , (Qk − Qk−1 )vj H 1 (Ω)
k=0 k=0 i,j=0
∞ min{i,j}

= (Qk − Qk−1 )vi , (Qk − Qk−1 )vj H 1 (Ω)
i,j=0 k=0
∞ min{i,j}

≤ (Qk − Qk−1 )vi H 1 (Ω) (Qk − Qk−1 )vj H 1 (Ω) .
i,j=0 k=0
By using the global inverse inequality (13.15), the stability of the L2 projection
(see Remark 9.14), and applying some interpolation argument, we obtain from
assumption (13.12) for the already fixed parameter s ∈ (0, 1] the estimate
(Qk − Qk−1 )vi H 1 (Ω) ≤ c h−s

k (Qk − Qk−1 )vi H 1−s (Ω)
≤ c h−s
k vi H 1−s (Ω) .
Moreover,
vi H 1−s (Ω) = (Q1i − Q1i−1 )vH 1−s (Ω)

= (Q1i − Q1i−1 )(Q1i − Q1i−1 )vH 1−s (Ω)
≤ (Q1i − I)vi H 1−s (Ω) + (I − Q1i−1 )vi H 1−s (Ω)
≤ c hsi vi H 1 (Ω) .
Hence we obtain
∞ ∞ min{i,j}

(Qk − Qk−1 )v2H 1 (Ω) ≤ c h−2s
k hsi hsj vi H 1 (Ω) vj H 1 (Ω) .
k=0 i,j=0 k=0
By using assumption (13.12) we further have

−k −2s 9 :−2s
h−2s
k ≤ c 2 = c 2min{i,j}−k
22s min{i,j} .
Then, for the already fixed parameter s ∈ (0, 1] it follows that

min{i,j} min{i,j}
min{i,j}−k
h−2s
k ≤ c 22s min{i,j} 2−2s c 22s min{i,j} .
≤
k=0 k=0
By using assumption (13.12) this gives

∞
∞

(Qk − Qk−1 )v2H 1 (Ω) ≤ c 22s min{i,j} 2−s(i+j) vi H 1 (Ω) vj H 1 (Ω)
k=0 i,j=0
∞
=c 2−s|i−j| vi H 1 (Ω) vj H 1 (Ω) .
i,j=0
If we define a symmetric matrix A by its entries A[j, i] = 2−s|i−j| , we then get

∞
∞

(Qk − Qk−1 )v2H 1 (Ω) ≤ c A2 vi 2H 1 (Ω) .
k=0 i=0
By using the estimate (13.18) of the Schur lemma we obtain

∞

A2 ≤ sup 2−s|i−j| .
j∈N0 i=0
For q := 2−s < 1 and for some j ∈ N0 this norm is bounded by

∞

j−1 ∞

j ∞
∞
2
q |i−j| = q j−i + q i−j = qi + qi ≤ 2 qi = ,
i=0 i=0 i=j i=1 i=0 i=0
1−q
and therefore it follows that

∞
∞

(Qk − Qk−1 )v2H 1 (Ω) ≤ c̄ vi 2H 1 (Ω) .
k=0 i=0
Finally,
∞
∞

vi 2H 1 (Ω) = (Q1i − Q1i−1 )v, (Q1i − Q1i−1 )vH 1 (Ω)
i=0 i=0
∞

= (Q1i − Q1i−1 )(Q1i − Q1i−1 )v, vH 1 (Ω)
i=0
∞

= (Q1i − Q1i−1 )v, vH 1 (Ω) = v, vH 1 (Ω) = v2H 1 (Ω) ,
i=0
which gives the upper estimate.

To prove the lower estimate we use the strengenthed Cauchy–Schwarz
inequality (Lemma 13.18) for some q < 1 to obtain
∞

||v||2H 1 (Ω) = (Qi − Qi−1 )v, (Qj − Qj−1 )vH 1 (Ω)
i,j=0
∞

≤ q |i−j| (Qi − Qi−1 )vH 1 (Ω) (Qj − Qj−1 )vH 1 (Ω) .
i,j=0
Now the assertion follows as above by applying the Schur lemma.

Remark 13.19. For s ∈ [0, 32 ) we may define the more general multilevel oper-
ator
∞
B s := h−2s
k (Qk − Qk−1 )
k=0
which satisfies, as in the special case s = 1, the spectral equivalence inequali-

ties
c1 v2H s (Ω) ≤ B s v, vL2 (Ω) ≤ c2 v2H s (Ω) for all v ∈ H s (Ω).

Although the following considerations are only done for the special case s = 1,
these investigations can be extended to the more general case s ∈ [0, 32 ), too.
By using Theorem 13.13 the multilevel operator B 1 : H 1 (Ω) → H −1 (Ω)
1
is bounded and H (Ω)–elliptic. The inverse operator (B ) 1 −1
: H (Ω) →
−1
1 −1
H (Ω) is then bounded and H (Ω)–elliptic, in particular the spectral equiv-
alence inequalities (13.13) are valid. For the inverse operator (B 1 )−1 again a
multilevel representation can be given.
Lemma 13.20. The inverse operator (B 1 )−1 allows the representation

∞

B −1 := (B 1 )−1 = h2k (Qk − Qk−1 ).
k=0
Proof. The assertion follows directly from

∞
∞
B −1 B 1 = h−2 2
k hj (Qk − Qk−1 )(Qj − Qj−1 )
k=0 j=0
∞
= (Qk − Qk−1 ) = I.

k=0
Remark 13.21. If we define the L2 projection operators Qj : L2 (Ω) → Sh0j (Ω)

onto the space of piecewise constant basis functions ϕ0,j k , then the related
multilevel operator B s satisfies the spectral equivalence inequalities
c1 v2H s (Ω) ≤ B s v, vL2 (Ω) ≤ c2 v2H s (Ω) for all v ∈ H s (Ω)
where s ∈ (− 12 , 12 ).
By using corollary 13.7 we can now establish the spectral equivalence of the
system matrix AhL with the discrete preconditioning matrix
A = M̄h B −1 M̄h
C L hL L
where
BhL [, k] = B −1 ϕL
k , ϕ L2 (Ω) ,
L
M̄hL [, k] = ϕL
k , ϕ L2 (Ω)
L
1
k , ϕ ∈ VL = ShL (Ω).
for all ϕL L
−1 r inside the algorithm of

It remains to describe the application v = CA
the preconditioned method of conjugate gradients, see Algorithm 13.3. There
we have to compute
−1 r = M −1 Bh M −1 r,
v := CA hL L hL
or,
u := Mh−1
L
r, w := BhL u, v := Mh−1
L
w.
By using the isomorphism u ∈ RML ↔ uhL ∈ VL we obtain for the components
of w = BhL u

ML
ML
w := BhL [, k]uk = B −1 ϕL
k , ϕ L2 (Ω) uk = B
L −1
L2 (Ω) .
uhL , ϕL
k=1 k=1
Hence, for uhL ∈ VL we need to evaluate

∞

L
zhL := B −1 uhL = h2k (Qk − Qk−1 )uhL = h2k (Qk − Qk−1 )uhL ∈ VL
k=0 k=0
which is a finite sum due to Qk uhL = uhL for k ≥ L. For the components of
w = BhL u we then obtain

ML
−1
w = B L2 (Ω)
uhL , ϕL = zhL , ϕL
L2 (Ω) = zk ϕL
k , ϕ L2 (Ω) .
L
k=1
w = MhL z,
and therefore there is no need to invert the inverse mass matrix MhL when
computing the preconditioned residual,
v = Mh−1
L
w = Mh−1
L
MhL z = z .
It remains to compute the coefficients of z ∈ RML ↔ zhL ∈ VL . For this we

have the representation

L
zhL = h2k (Qk − Qk−1 )uhL
k=0

L−1
= h2L QL uhL + (h2k − h2k+1 )Qk uhL
k=0

L−1
= h2L ūhL + (h2k − h2k+1 )ūhk
k=0
where

Mk
ūhk = Qk uhL = ūk ϕk ∈ Vk
=1
is the L2 projection of uhL into the trial space Vk , k = 0, 1, . . . , L. Due to

c h2k ≤ h2k − h2k+1 ≤ h2k
we can define

L
z̄hL := h2k ūhk
k=0
to be spectrally equivalent to C −1 . The evaluation of z̄h can be done recur-

A L
2
sively. Starting from z̄h0 = h0 ūh0 ∈ V0 we have

Mk−1

Mk
z̄hk := z̄hk−1 + h2k ūhk = z̄ k−1 ϕk−1
+ h2k ūk ϕk .
=1 =1
Due to the inclusion Vk−1 ⊂ Vk we can write each basis function ϕk−1
∈ Vk−1
as a linear combination of basis functions ϕkj ∈ Vk ,

Mk
ϕk−1
= k
r ,j ϕkj for all = 1, . . . , Mk−1 .
j=1
Hence we can write

Mk−1

Mk−1

Mk
Mk M k−1
z̄ k−1 ϕk−1
= z̄ k−1 k
r ,j ϕkj = z̄ k−1 r ,j
k
ϕkj .
=1 =1 j=1 j=1 =1
By introducing the matrices

k
Rk−1,k [j, ] = r ,j for j = 1, . . . , Mk , = 1, . . . , Mk−1
we obtain for the coefficient vector
z̄ k := Rk−1,k z̄ k−1 + h2k ūk .

k
When considering a uniform mesh refinement strategy the coefficients r ,j are
k−1
given by the nodal interpolation of the basis functions ϕ ∈ Vk−1 at the
nodes xj of the decomposition TNk , see Fig. 13.1.
By using the matrices
Rk := RL−1,L . . . Rk,k+1 for k = 0, . . . , L − 1, RL := I
we obtain by induction

L
z̄ L = h2k Rk ūk .
k=0
It remains to compute the L2 projections ūhk = Qk uhL as the unique solutions

of the variational problems
ūhk , ϕk L2 (Ω) = uhL , ϕk L2 (Ω) for all ϕk ∈ Vk .

0 0 0 0 0
@
@ @
@ 1 @
@ 1
@ @2 @
2
@ @ @
@ @ @0
@ 0 @
1 @ 1
@ @ @
@ @1 @1
@ 2
@
2@
@ @ @
0 0 0 0 0
Fig. 13.1. Basis functions ϕk−1
and coefficients k
r,j (d = 2).
This is equivalent to a linear system of algebraic equations,
Mhk ūk = f k
where
Mhk [, j] = ϕkj , ϕk L2 (Ω) , f k = uhL , ϕk L2 (Ω) .
In particular for k = L we have
f L = MhL u = MhL Mh−1

L
r = r
and therefore
ūL = Mh−1
L
r.
Due to

Mk
M
f k−1 = uhL , ϕk−1
L2 (Ω) = k
r ,j uhL , ϕkj L2 (Ω) = k
r ,j fjk
j=1 j=1
we get

f k−1 = Rk−1,k f k = Rk−1 r.
By recursion we therefore have
ūk = Mh−1
k
Rk r,
and the application of the preconditioner reads

L
v = h2k Rk Mh−1
k
Rk r .
k=0
Taking into account the spectral equivalence of the mass matrices with the
diagonal matrices hdk I, see Lemma 9.7, we then obtain for the application of
the multilevel preconditioner

L
v = h2−d
k Rk Rk r. (13.19)
k=0
13.3 Solution Methods for Saddle Point Problems 319
The realization of the multilevel preconditioner (13.19) therefore requires the

restriction of a residual vector r which is given on the computational level
L, and a weighted summation of prolongated coarse grid vectors. Thus, an
application of the multilevel preconditioner requires O(M ) operations only.
CA = I CA = M̄hL Bh−1
L
M̄hL
−1 −1
L M λmin λmax κ(CA Ah ) λmin λmax κ(CA Ah )
1 13 2.88 –1 6.65 23.13 16.34 130.33 7.98
2 41 8.79 –2 7.54 85.71 16.69 160.04 9.59
3 145 2.42 –2 7.87 324.90 16.32 179.78 11.02
4 545 6.34 –3 7.96 1255.75 15.47 193.36 12.50
5 2113 1.62 –3 7.99 4925.47 15.48 202.94 13.11
6 8321 4.10 –4 8.00 19496.15 15.58 209.85 13.47
7 33025 ≈80000 15.76 214.87 13.63
8 131585 ≈320000 15.87 218.78 13.79
9 525313 ≈1280000 15.96 221.65 13.89
Theory: O(h−2 ) O(1)
Table 13.3. Extremal eigenvalues and spectral condition number (FEM).
In Table 13.3 we give the extremal eigenvalue and the resulting spec-
tral condition numbers of the preconditioned finite element stiffness matrix
−1
CA [a a +AhL ]. This preconditioner is also needed for an efficient solve of the
linear system (11.22), as it will be considered in the next section. The results
for the non–preconditioned system (CA = I) confirm the statement of Lemma
11.4, while the boundedness of the spectral condition of the preconditioned
systems coincides with the results of this section.
13.3 Solution Methods for Saddle Point Problems

The boundary element discretization of the symmetric formulation of bound-
ary integral equations to solve mixed boundary value problems, as well as
the finite element discretization of saddle point problems, both lead to linear
systems of algebraic equations of the form
& '& ' & '
A −B u1 f1
= (13.20)
B D u2 f2
where the block A ∈ RM1 ×M1 is symmetric and positive definite, and
where D ∈ RM2 ×M2 is symmetric but positive semi–definite. Accordingly,
B ∈ RM1 ×M2 . Since the matrix A is assumed to be positive definite, we can
solve the first equation in (13.20) for u1 to obtain
u1 = A−1 Bu2 + A−1 f 1 .
Inserting this into the second equation of (13.20) this results in the Schur
complement system
0 1
D + B A−1 B u2 = f 2 − B A−1 f 1 (13.21)
where
S = D + B A−1 B ∈ RM2 ×M2 . (13.22)
is the Schur complement. From the symmetry properties of the block matrices
A, B and D we conclude the symmetry of S, while, at this point, we assume
the positive definiteness of S.
We assume that for the symmetric and positive definite matrices A and
S = D + B A−1 B there are given some positive definite and symmetric
preconditioning matrices CA and CS satisfying the spectral equivalence in-
equalities
cA1 (CA x1 , x1 ) ≤ (Ax1 , x1 ) ≤ c2 (CA x1 , x1 )
A
(13.23)
for all x1 ∈ RM1 as well as
cS1 (CS x2 , x2 ) ≤ (Sx2 , x2 ) ≤ cS2 (CS x2 , x2 ) (13.24)
for all x2 ∈ RM2 . Hence, to solve the Schur complement system (13.21) we
can apply the CS preconditioned method of conjugate gradients (Algorithm
13.3). There, the matrix by vector multiplication sk = Spk for the Schur
complement (13.22) reads
sk := Dpk + B A−1 Bpk = Dpk + B wk ,
where wk is the unique solution of the linear system
Awk = Bpk .
This system can be solved either by a direct method, for example by the
Cholesky approach, or again by using a CA preconditioned method of con-
jugate gradients (Algorithm 13.3). Depending on the application under con-
sideration the Schur complement approach can be disadvantageous. Then, an
iterative solution strategy for the system (13.20) should be used. Possible iter-
ative solution methods for general non–symmetric linear systems of the form
(13.20) are the method of the generalized minimal residual (GMRES, [120]),
or the stabilized method of biorthogonal search directions (BiCGStab, [155]).
Here, following [26], we will describe a transformation of the block–skew
symmetric but positive definite system (13.20) leading to a symmetric and
positive definite system for which a preconditioned conjugate gradient ap-
proach can be used.
For the preconditioning matrix CA we need to assume that the spectral equiv-
alence inequalities (13.23) hold where
cA
1 > 1 (13.25)
is satisfied. This can always be guaranteed by using an appropriate scaling,

i.e. for a given preconditioning matrix CA we need to compute the minimal
−1
eigenvalue of the preconditioned system CA A. From the assumption (13.25)
we find that the matrix A − CA is positive definite,
((A − CA )x1 , x1 ) ≥ (cA

1 − 1) (CA x1 , x1 ) for all x1 ∈ R
M1
,
and hence invertible. Thus, also the matrix

−1 −1
ACA − I = (A − CA )CA
is invertible, and & '

−1
ACA −I 0
T = −1
−B CA I
defines a invertible matrix. By multiplying the linear system (13.20) with the
transformation matrix T this gives
& '& '& ' & '& '
−1 −1
ACA −I 0 A −B u1 ACA −I 0 f1
−1
= −1 (13.26)
−B CA I B D u2 −B CA I f2
where the system matrix

& '& '
−1
ACA −I 0 A −B
M = −1
−B CA I B D
& '
−1 −1
ACA A−A (I − ACA )B
= −1 −1 (13.27)
B (I − CA A) D + B CA B
is symmetric. From the spectral equivalence inequalities of the transformed

system matrix M with the preconditioning matrix
& '
A − CA 0
CM := (13.28)
0 CS
then the positive definiteness of M follows. Hence we can use a preconditioned

conjugate gradient scheme to solve the transformed linear system (13.26).
Theorem 13.22. For the preconditioning matrix CM as defined in (13.28)

there hold the spectral equivalence inequalities
1 (CM x, x) ≤ (M x, x) ≤ c2 (CM x, x)
cM M
for all x ∈ RM1 +M2
where
8
1 A 1 A
cM
1 = c2 [1 + c1 ] −
S
[c (1 + cS1 )]2 − cS1 cA
2 ,
2 4 2
8
1 1 A
cM
2 = cA [1 + cS2 ] + [c (1 + cS2 )]2 − cS2 cA
2 .
2 2 4 2
Proof. We need to estimate the extremal eigenvalues of the preconditioned
−1
system matrix CM M , in particular we have to consider the eigenvalue prob-
lem
& '& ' & '& '
−1 −1
ACA A − A (I − ACA )B x1 A − CA 0 x1
−1 −1 = λ .
B (I − CA A) D + B CA B x2 0 CS x2
Let λi be an eigenvalue with associated eigenvectors x1i and xi2 . From the first
equation,
−1 −1
(ACA A − A)xi1 + (I − ACA )Bxi2 = λi (A − CA )xi1 ,
we find by some simple manipulations
−Bxi2 = (λi CA − A)xi1 .
For λi ∈ [1, cA
2 ] nothing is to be shown. Hence we only consider λi ∈ [1, c2 ]
A
were λi CA − A is invertible. Thus,
xi1 = −(λi CA − A)−1 Bxi2 .
Inserting this result into the second equation of the eigenvalue problem,
−1 −1
B (I − CA A)xi1 + [D + B CA B]xi2 = λi CS xi2 ,
this gives
−1 −1
−B CA (CA − A)(λi CA − A)−1 Bxi2 + [D + B CA B]xi2 = λi CS xi2 .
Due to
−1 −1
−CA (CA − A)(λi CA − A)−1 = −CA [λi CA − A + (1 − λi )CA ](λi CA − A)−1
−1
= (λi − 1)(λi CA − A)−1 − CA
this is equivalent to
(λi − 1)B (λi CA − A)−1 Bxi2 + Dxi2 = λi CS xi2 .
When λi > cA 2 is satisfied we have that λi CA − A is positive definite. By using

the spectral equivalence inequalities (13.23) we then obtain
λi − cA
2
(Ax1 , x1 ) ≤ ((λi CA − A)x1 , x1 )
cA
2
for all x1 ∈ RM1 , and therefore
cA
2
((λi CA − A)−1 x1 , x1 ) ≤ (A−1 x1 , x1 ) for all x1 ∈ Rn1 .
λi − cA
2
From the spectral equivalence inequalities (13.24) we conclude
λi
(Sxi2 , xi2 ) ≤ λi (CS xi2 , xi2 )
cS2
= (Dxi2 , xi2 ) + (λi − 1)((λi CA − A)−1 Bxi2 , Bxi2 )
cA
2
≤ (Dxi2 , xi2 ) + (λi − 1) (A−1 Bxi2 , Bxi2 )
λi − cA
2
A λi − 1
≤ c2 i i
(Sx2 , x2 )
λi − cA 2
and therefore
λi λi − 1
≤ cA
2 ,
S
c2 λi − cA
2
i.e.
λ2i − cA
2 [1 + c2 ]λi + c2 c2 ≤ 0 .
S S A
From this we obtain

λ− ≤ λi ≤ λ+
where 8
1 A 1 A
λ± = c2 [1 + c2 ] ±
S
[c (1 + cS2 )]2 − cA S
2 c2 .
2 4 2
8
1 A 1 A
c2 < λi ≤ c2 [1 + c2 ] +
A S
[c (1 + cS2 )]2 − cA S M
2 c2 = c2 .
2 4 2
It remains to consider the case λi < 1 where A − λi CA is positive definite. By
using the spectral equivalence inequalities (13.23) we get
2 − λi
cA
((A − λi CA )x1 , x1 ) ≤ (Ax1 , x1 )
cA
2
and therefore
cA
2
((A − λi CA )−1 x1 , x1 ) ≥ (A−1 x1 , x1 )
2 − λi
cA
for all x1 ∈ RM1 . Again, by using the spectral equivalence inequalities (13.24)
we conclude
λi
(Sxi2 , xi2 ) ≥ λi (CS xi2 , xi2 )
cS1
= (Dxi2 , xi2 ) + (1 − λi )((A − λi CA )−1 Bxi2 , Bxi2 )
cA
≥ (Dxi2 , xi2 ) + (1 − λi ) A 2 (A−1 Bxi2 , Bxi2 )
c2 − λi
cA
2
≥ (1 − λi ) (Sxi2 , xi2 )
2 − λi
cA
and therefore
1 − λi λi
cA
2 ≤ S.
c2 − λi
A c1
λ2i − cA
2 [c1 + 1]λi + c1 c2 ≤ 0,
S S A
i.e.
λ− ≤ λi ≤ λ+
where 8
1 A 1 A
λ± = c2 [1 + c1 ] ±
S
[c (1 + cS1 )]2 − cS1 cA
2.
2 4 2
Summarizing we have
8
1 A 1 A
1 > λi ≥ c2 [1 + c1 ] −
S
[c (1 + cS1 )]2 − cS1 cA M
2 = c1 .
2 4 2
This completes the proof.
For the solution of the transformed linear system (13.26) Algorithm
13.3 of the preconditioned conjugate gradient approach can be applied. On
a first glance the multiplication with the inverse preconditioning matrix
−1 k+1
v k+1 = CM r , in particular the evaluation of v k+1
1 = (A−CA )−1 rk+1 seems
to be difficult. However, from the recursion of the residual,
rk+1 = rk − αk M pk , we find the representation
−1
rk+1
1 := rk1 − αk (ACA − I)(Apk1 − Bpk2 ).
Hence we can write the preconditioned residual v k1 recursively as

−1
v k+1
1 := v k1 − αk CA (Apk1 − Bpk2 ).
In particular for k = 0 we have

2 3
−1
v 01 := CA Ax01 − Bx02 − f 1 .
The resulting preconditioned iterative scheme is summarized in Algorithm

13.4.
For an arbitrary initial guess u0 ∈ RM1 +M2 compute the residual

r̄01 := Au01 − Bu02 − f 1 , r̄02 := B u01 + Du02 − f 2 .
Compute the transformed residual
−1 0
w01 := CA r̄1 , r01 := Aw01 − r̄01 , r02 := r̄02 − B w01 .
Initialize the method of conjugate gradients:
v 01 := w01 , v 02 := CS−1 r02 , p0 := v 0 ,
0 := (v 0 , r0 ).
For k = 0, 1, 2, . . . , n − 1:
Realize the matrix by vector multiplication
sk1 := Apk1 − Bpk2 , sk2 := B pk1 + Dpk2 .
Compute the transformation
−1 k
wk1 := CA s1 , sk1 := Awk1 − sk1 , sk2 := sk2 − B wk1 .
Compute the new iterates
σk := (sk , pk ), αk :=
k /σk ;
uk+1 := uk − αk pk , rk+1 := rk − αk sk ;
v k+1
1 := v k1 − αk wk+1
1 , v k+1
2 := CS−1 rk+1
2 ,
k+1 := (v k+1 , rk+1 ) .
Stop, if
k+1 ≤ ε
Otherwise compute the new search direction
βk :=
k+1 /
k , pk+1 := v k+1 + βk pk .
Algorithm 13.4: Conjugate gradient method with Bramble/Pasciak transformation.
L N M Schur CG BP CG
2 16 11 11 16
3 32 23 13 19
4 64 47 14 21
5 128 95 14 21
6 256 191 15 23
7 512 383 16 23
8 1024 767 16 23
9 2048 1535 16 24
Table 13.4. Comparison of Schur CG and Bramble/Pasciak CG.
As an example we consider the solution of the linear system (12.43) which

results from a Galerkin boundary element approximation, see Section 12.3,
& '& ' & '
Vh − 12 Mh − Kh w 0
1
= . (13.29)
2 M h + K h Dh û f
The associated Schur complement system reads


1 −1 1
Sh û = Dh + ( Mh + Kh )Vh ( Mh + Kh ) û = f . (13.30)
2 2
As preconditioner for the Schur complement matrix Sh we can apply the

preconditioning strategy as described in section 13.2.1. But in this case the
spectral equivalence inequalities (13.11) of the hypersingular boundary in-
tegral operator D : H 1/2 (ΓN ) → H −1/2 (Γ ) and of the inverse single layer
potential V : H −1/2
(ΓN ) → H 1/2 (ΓN ) are not satisfied due to the differ-
ent function spaces to be used. But for finite dimensional conformal trial
spaces Sh1 (ΓN ) ⊂ H 1/2 (ΓN ) one can prove related estimates [104], i.e. for all
1
vh ∈ Sh (ΓN ) there hold the spectral equivalence inequalities
γ1 V −1 vh , vh Γ ≤ Dvh , vh Γ ≤ γ2 [1 + log |h|]2 V −1 vh , vh Γ .
When using the preconditioning matrix CD = M̄h V̄h−1 M̄h we then obtain the
estimate for the spectral condition number,
−1
κ2 (CD Dh ) ≤ c [1 + log |h|]2 .
As described in section 13.2.1 we can also define a preconditioning matrix CV

for the discrete single layer potential Vh which is based on the modified hyper-
singular boundary integral operator D̂ : H 1/2 (Γ ) → H −1/2 (Γ ), see [144]. In
Table 13.4 we give the number of iterations of the preconditioned conjugate
gradient approach for the solution of the Schur complement system (13.30),
and of the conjugate gradient approach with the Bramble/Pasciak transfor-
mation to solve the system (13.29). As relative accuracy we have considered
ε = 10−8 , and the scaling of the preconditioning matrix CV of the discrete sin-
gle layer potential was chosen such that the spectral equivalence inequalities
(13.23) are satisfied with cA 1 = 1.2.
14
Fast Boundary Element Methods
Boundary element methods as described in Chapter 12 result in dense stiff-

ness matrices. In particular, both the storage requirements and the numerical
amount of work to compute all entries of a boundary element stiffness matrix
is quadratic in the number of degrees of freedom. Hence there is a serious
need to derive and to describe fast boundary element methods which exhibit
an almost linear, up to some polylogarithmic factors, behavior in the num-
ber of degrees of freedom. Here we constrict our considerations to the case
of a two–dimensional model problem, for the three–dimensional case, see, for
example, [117].
As a model problem we consider the Dirichlet boundary value problem
−∆u(x) = 0 for x ∈ Ω ⊂ R2 , γ0int u(x) = g(x) for x ∈ Γ = ∂Ω
where we assume diam Ω < 1 to ensure the invertibility of the single layer
potential V . When using an indirect single layer potential (7.4) the solution
of the above problem is given by

1
x) = −
u( log |
x − y|w(y)dsy for x ∈ Ω.
2π
Γ
The yet unknown density w ∈ H −1/2 (Γ ) is then given as the unique solution
of the boundary integral equation (7.12),

1
(V w)(x) = − log |x − y|w(y)dsy = g(x) for x ∈ Γ.
2π
Γ
Let Sh0 (Γ ) = span{ϕ0k }N

k=1 be the trial space of piecewise constant basis func-
tions ϕ0k which are defined with respect to a globally quasi–uniform boundary
mesh {τk }N k=1 with a global mesh size h. Then we can find an approxi-
mate solution wh ∈ Sh0 (Γ ) as the unique solution of the Galerkin variational
problem
328 14 Fast Boundary Element Methods
V wh , τh Γ = g, τh Γ for all τh ∈ Sh0 (Γ ). (14.1)

This variational problem is equivalent to a linear system Vh w = f of algebraic
equations where the stiffness matrix Vh is defined as

1
Vh [, k] = V ϕ0k , ϕ0 Γ = − log |x − y|dsy dsx (14.2)
2π
τ τk
for all k, = 1, . . . , N . Note that for the approximate solution wh ∈ Sh0 (Γ )

w − wh H −1/2 (Γ ) ≤ c h3/2 |w|Hpw

1 (Γ ) . (14.3)
1
when assuming w ∈ Hpw (Γ ). Due to the nonlocal definition of the fundamen-
tal solution the stiffness matrix Vh is dense, i.e. to describe the symmetric
matrix Vh we need to store 12 N (N + 1) matrix entries. Moreover, a realization
of a matrix by vector product within the use of a preconditioned conjugate
gradient scheme (Algorithm 13.3) requires N 2 multiplications. Hence we have
a quadratic amount of work in both storage of the matrix and in a matrix by
vector product with respect to the number N of degrees of freedom. In con-
trast to standard boundary element methods we are interested in the design
of fast boundary element methods where the numerical amount of work will
be of the order O(N (log2 N )α ) where we have to ensure an error estimate as
given in (14.3) for a standard boundary element method.
In this chapter we will consider two different approaches to derive fast
boundary element methods. The use of wavelets [46, 47, 94, 127] leads to dense
stiffness matrices Vh , but since most of the matrix entries can be neglected this
results in a sparse approximation Vh of the stiffness matrix. A second approach
is based on a hierarchical clustering of boundary elements [14, 62, 65, 73, 122]
which defines a block partitioning of the stiffness matrix Vh . If two clusters
are well separated the related block can be approximated by some low rank
matrices.
14.1 Hierarchical Cluster Methods

1
Since the fundamental solution U ∗ (x, y) = − 2π log |x − y| is only a function
of the distance |x − y|, all matrix entries Vh [, k] of the discrete single layer
potential Vh as defined in (14.2) only depend on the distance, on the size, and
on the shape of the boundary elements τk and τ . Hence we can cluster all
boundary elements when taking into account their size and the distances to
each other. The ratio of the cluster size and the distance between two clusters
will then serve as an admissibility criterion to define an approximation of the
fundamental solution. A larger distance between two clusters then also allows
to consider larger clusters of boundary elements. For this we will consider
an appropriate hierarchy of clusters. The interaction between two boundary
14.1 Hierarchical Cluster Methods 329
elements τk and τ to compute the matrix entry Vh [, k] is then replaced by

the interaction between the associated clusters. <N
For a given globally quasi–uniform boundary decomposition Γ = =1 τ
we first have to construct a suitable cluster hierarchy. Due to the general
assumption diam Ω < 1 we may assume Ω ⊂ (0, 1)2 . Hence, all boundary ele-
ments τ are contained in the surrounding square box Ω10 = (0, 1)2 . Applying
a recursive decomposition of Ωjλ−1 into four congruent boxes Ωiλ as depicted
in Fig. 14.1 this first defines a hierarchy of boxes Ωiλ , and from this we easily
find a hierarchical clustering of the boundary elements τ . The number λ of
recursive refinement steps is called the level of the hierarchy.
2 2 2 2
Ω11 Ω12 Ω15 Ω16
Ω31 Ω41
Ω92 2
Ω10 2
Ω13 2
Ω14
Ω10
Ω32 Ω42 Ω72 Ω82
Ω11 Ω21
Ω12 Ω22 Ω52 Ω62
Fig. 14.1. Hierarchy of boxes Ωjλ for λ = 0, 1, 2.
For λ = 1, . . . , L we therefore have the representation

4j
*
λ−1 λ
Ωj = Ωi , j = 1, . . . , 4(λ−1) (14.4)
i=4(j−1)+1
where the length dλj of an edge of the box Ωjλ is given by
dλj = 2−λ , j = 1, . . . , 4λ .
The refinement strategy (14.4) is applied recursively until the edge length dL
j
of a box ΩjL on the finest level L is proportional to the mesh size h of the
globally quasi–uniform boundary mesh {τ }N =1 , i.e.
−L
dL
j = 2 ≤ cL h
induces a maximal number of boundary elements τ which are contained in
the box ΩjL . Then we find for the maximal level of the cluster tree
cL ln(1/h)
L ≥ .
ln 2
Since the boundary decomposition is assumed to be globally quasi–uniform,
this implies that the surface measure |Γ | is proportional to N h and therefore
we obtain
L = O(ln N ) . (14.5)
To describe the clustering of the boundary elements {τ }N
we may consider
=1
the clustering of the associated element midpoints x̂ ∈ τ for = 1, . . . , N .
For j = 1, . . . , 4L we first collect all boundary elements τ where the midpoint
x̂ is in the box ΩjL in a cluster ωjL ,
*
ωLj := τ .
x̂ ∈ΩjL
The hierarchy (14.4) of boxes Ωjλ now transfers directly to a hierarchy of the
related clusters ωjλ , see Fig. 14.2,
4j
*
ω λ−1
j := ω λi for j = 1, . . . , 4(λ−1) , λ = L, . . . , 1. (14.6)
i=4(j−1)+1
ω10
P
@PPP
@ P
P
ω11 ω21 ω31 ω41
··
··
··
··
ω1L−2
P
@PPP
@ P
P
ω1L−1 ω2L−1 ω3L−1 ω4L−1
P
@PPP
@ P
P
ω1L ω2L ω3L ω4L
A
@
A@
τ1 · · · · · · τ1
Fig. 14.2. Cluster tree ωjλ .
For each cluster ωjλ we define

Ijλ := ∈ N : τ ⊂ ωjλ
as the index set of all associated boundary elements τ where
Pjλ : I10 = {1, 2, . . . , N } → Ijλ
describes the assignment of the boundary elements {τ }N

=1 to the associated
cluster ωjλ . Finally,
Njλ := dim ωjλ
14.1 Hierarchical Cluster Methods 331
denotes the number of boundary elements τ inside the cluster ωjλ . By con-
struction we have for each λ = 0, 1, 2, . . . , L
4

λ
Njλ = N . (14.7)
j=1
By
diam ωjλ := sup |x − y|
x,y∈ωjλ
we denote the diameter of the cluster ωjλ , and by
dist(ωiκ , ωjλ ) := inf |x − y|

(x,y)∈ωiκ ×ωjλ
we define the distance between the clusters ωiκ and ωjλ . A pair of clusters ωiκ
and ωjλ is called admissible, if

dist(ωiκ , ωjλ ) ≥ η max diam ωiκ , diam ωjλ (14.8)
is satisfied where η > 1 is a prescribed parameter. For simplicity we only

consider admissible clusters ωiκ and ωjλ which are defined for the same level
κ = λ. If there are given two admissible clusters ωiλ and ωjλ , then also all
subsets ωiλ+1
⊂ ωiλ and ωjλ+1
⊂ ωjλ are admissible. Therefore, a pair of clusters
ωiλ and ωjλ is called maximally admissible if there exist inadmissible clusters
ωiλ−1
and ωjλ−1
where ωiλ ⊂ ωiλ−1
and ωjλ ⊂ ωjλ−1
are satisfied.
For the stiffness matrix Vh as defined in (14.2) the hierarchical clustering
(14.6) allows the representation

L 4

λ
4 λ
4

L
4 L
Vh = (Pjλ ) Vhλ,ij Piλ + (PjL ) VhL,ij PiL (14.9)

λ=0 j=1 i=1 j=1 i=1
4 56 7 4 56 7
ωiλ ,ωjλ maximally admissible ωiL ,ωjL inadmissible
where the block matrices Vhλ,ij ∈ RNj ×Ni , are defined by

λ λ

λ,ij 1
Vh [, k] = − log |x − y|dsy dsx for τk ∈ ωiλ , τ ∈ ωjλ , (14.10)
2π
τ τk
see also Fig. 14.3. The sum is to be taken over all inadmissible clusters ωiL and
ωjL includes in particular the interaction of a cluster with itself and with all
neighboring clusters. Hence we denote this part as the near field of the stiffness
matrix Vh while the remainder, i.e. the sum over all maximally admissible
clusters is called the far field. A box ΩiL has maximal 8 direct neighbors
ΩjL , the associated cluster ωiL therefore has a certain number of inadmissible
Fig. 14.3. Hierarchical partitioning of the stiffness matrix Vh .
clusters ωjL where the number only depends on the parameter η > 1. All other
clusters are thus admissible and therefore they are included in the far field.
The near field part therefore contains only O(η 2 4L ) = O(η 2 N ) summands.
While the block matrices of the near field part can be evaluated directly as
in a standard boundary element method, the block matrices of the far field
part can be approximated by using low rank matrices which allow for a more
efficient application. The resulting matrices are called hierarchical matrices,
or H matrices. [72].
14.2 Approximation of the Stiffness Matrix

For a maximally admissible pair of clusters ωiλ and ωjλ we have to compute
all entries of the block matrix Vhλ,ij ,

Vhλ,ij [, k] = U ∗ (x, y)dsy dsx for τk ∈ ωiλ , τ ∈ ωjλ .
τ τk
The basic idea for the derivation of fast boundary element methods is an
1
approximate splitting of the fundamental solution U ∗ (x, y) = − 2π log |x − y|
into functions which only depend on the integration point y ∈ ωi , and on the
λ
observation point x ∈ ωjλ ,

14.2 Approximation of the Stiffness Matrix 333

U∗ (x, y) = λ,j
fm λ,i
(x)gm (y) for (x, y) ∈ ωjλ × ωiλ . (14.11)
m=0
We assume that there is given an error estimate
|U ∗ (x, y) − U∗ (x, y)| ≤ c(η, ) for (x, y) ∈ ωjλ × ωiλ , (14.12)
where c(η, ) is a constant which only depends on the admissibility parameter

η and on the approximation order . By using the decomposition (14.11) we
can approximate the entries of the block matrix Vhλ,ij by computing

λ,ij
Vh [, k] := U∗ (x, y)dsy dsx for τk ∈ ωiλ , τ ∈ ωjλ . (14.13)
τ τk
From the error estimate (14.12) we then obtain
|Vhλ,ij [, k] − Vhλ,ij [, k]| ≤ c(η, ) ∆k ∆ for τk ∈ ωiλ , τ ∈ ωjλ . (14.14)
Inserting the series expansion (14.11) into (14.13) this gives

Vhλ,ij [, k] = λ,j
fm (x)dsx λ,i
gm (y)dsy for all τk ∈ ωiλ , τ ∈ ωjλ .
m=0 τ τk

Hence, for all boundary elements τk ∈ ωiλ and τ ∈ ωjλ we need to compute
vectors defined by the entries

aλ,j
m, := f λ,j
m (x)ds x , bλ,i
m,k := λ,i
gm (y)dsy ,
τ τk
where = 1, . . . , Njλ , k = 1, . . . , Niλ and m = 0, . . . , . The numerical amount

of work to store and to apply the approximate block matrix

Vhλ,ij = aλ,j λ,i
m (bm ) ,
m=0
which is a matrix of rank + 1, is therefore
( + 1)(Niλ + Njλ ).
As in (14.9) we can now define an approximation Vh of the global stiffness

matrix Vh ,

L 4

λ
4 λ
4

L
4 L
Vh = (Pjλ ) Vhλ,ij Piλ + (PjL ) VhL,ij PiL . (14.15)

λ=0 j=1 i=1 j=1 i=1
4 56 7 4 56 7
ωiλ ,ωjλ maximally admissible ωiL ,ωjL inadmissible
Due to (14.7) the total amount of work to store and to apply the approximate
stiffness matrix Vh is proportional to, by taking into account the near field
part,
( + 1)(η + 1)2 (L + 1)N + η 2 N. (14.16)
Instead of the original linear system Vh w = f we now have to solve the per-
turbed system Vh w
= f with an associated approximate solution w h ∈ Sh0 (Γ ).
The stability and error analysis of the perturbed problem is based on the
Strang lemma (Theorem 8.3). Hence we need to prove the positive definiteness
of the approximate stiffness matrix Vh , which will follow from an estimate of
the approximation error Vh − Vh .
Lemma 14.1. For a pair of maximally admissible clusters ωiλ and ωjλ the
error estimate (14.12) is assumed. Let Vh be the approximate stiffness matrix
as defined in (14.15). Then there holds the error estimate
|((Vh − Vh )w, v)| ≤ c(η, ) |Γ | wh L2 (Γ ) vh L2 (Γ )
for all w, v ∈ RN ↔ wh , vh ∈ Sh0 (Γ ).
Proof. By using the error estimate (14.14) we first have

L 4
4 λ λ

|((Vh − Vh )w, v)| ≤ |Vhλ,ij [, k] − Vhλ,ij [, k]| |wk | |v |
λ=0 j=1 i=1 τk ∈ωiλ τ ∈ωjλ
4 56 7
ωiλ ,ωjλ maximally admissible

L 4
4
λ λ

≤ c(η, ) ∆k |wk | ∆ |v | .
λ=0 j=1 i=1 τk ∈ωiλ τ ∈ωjλ
4 56 7
ωiλ ,ωjλ maximally admissible
Due to the assumption of the maximal admissibility of clusters ωiλ and ωjλ
each pair of boundary elements τk and τ appears maximal only once. Hence
we have

N
N
|((Vh − Vh )u, v)| ≤ c(η, ) ∆k |wk | ∆ |v | .
k=1 =1
By applying the Hölder inequality we finally obtain

& '1/2 & '1/2

N
N
N
∆k |wk | ≤ ∆k ∆k wk2 = |Γ |1/2 wh L2 (Γ )
k=1 k=1 k=1
from which the assertion follows.

By using the error estimate of Lemma 14.1, by applying the inverse inequality
in Sh0 (Γ ), and by an appropriate choice of the parameter η for the definition of
the near field and of the approximation order we now can prove the positive
definiteness of the approximate stiffness matrix Vh .
Theorem 14.2. Let all assumptions of Lemma 14.1 be satisfied. For an ap-
propriate choice of the parameter η and the approximate stiffness matrix Vh
is positive definite, i.e.
1 V
(Vh w, w) ≥ c wh 2H −1/2 (Γ ) for all wh ∈ Sh0 (Γ ).
2 1
Proof. By using Lemma 14.1, the H −1/2 (Γ )–ellipticity of the single layer po-
tential V , and the inverse inequality in Sh0 (Γ ) we obtain
(Vh w, w) = (Vh w, w) + ((Vh − Vh )w, w)

≥ V wh , wh Γ − |((Vh − Vh )w, w)|
≥ cV1 wh 2H −1/2 (Γ ) − c(η, ) |Γ | wh 2L2 (Γ )
0 1
≥ cV1 − c(η, ) |Γ | c2I h−1 wh 2H −1/2 (Γ )
1 V
≥ c wh 2H −1/2 (Γ ) ,
2 1
if
cV1
c(η, ) ≤ h (14.17)
2|Γ |c2I
is satisfied.
In the same way as in the proof of Theorem 14.2 the boundedness of the
approximate single layer potential Vh follows. Due to the positive definiteness
of the approximate stiffness matrix Vh we then obtain the unique solvability
of the perturbed linear system Vh w = f . Moreover, as in Theorem 8.3 we
can also estimate the error w − w h H −1/2 (Γ ) of the computed approximate
solution wh ∈ Sh0 (Γ ).
Theorem 14.3. Let the parameter η and be chosen such that the approxi-
mate stiffness matrix Vh as defined in (14.15) is positive definite. The uniquely
determined solution w ∈ RN ↔ w h ∈ Sh0 (Γ ) of the perturbed linear system
Vh w
= f then satisfies the error estimate
c(η, ) h−1/2 wL2 (Γ ) .
h H −1/2 (Γ ) ≤ w − wh H −1/2 (Γ ) +
w − w
∈ RN be the uniquely determined solutions of the linear sys-

Proof. Let w, w
tems Vh w = f and Vh w
= f , respectively. Then there holds the orthogonality
relation
(Vh w − Vh w,
v) = 0 for all v ∈ RN .
Since the approximated stiffness matrix Vh is positive definite, we obtain, by

using Lemma 14.1
1 V
h 2H −1/2 (Γ ) ≤ (Vh (w − w),
c wh − w w − w)
2 1
= ((Vh − Vh )w, w − w)

≤ c(η, ) wh L2 (Γ ) wh − w
h L2 (Γ ) .
Applying Lemma 12.2 this gives the stability estimate
wh L2 (Γ ) ≤ wL2 (Γ ) + w − wh L2 (Γ ) ≤ c wL2 (Γ ) .
Then, using the inverse inequality in Sh0 (Γ ) we conclude
wh − w c(η, ) h−1/2 wL2 (Γ ) .
h H −1/2 (Γ ) ≤
The assigned error estimate now follows from the triangle inequality.
1
From the error estimate (14.3) we conclude, when assuming w ∈ Hpw (Γ ),
h H −1/2 (Γ ) ≤ c1 h3/2 |w|Hpw

w − w c(η, ) h−1/2 wL2 (Γ ) .
1 (Γ ) +
To ensure an asymptotically optimal order of convergence we therefore need

to satisfy the condition
c(η, ) ≤ c2 h2 .
(14.18)
In this case, the error estimate
h H −1/2 (Γ ) ≤ c1 h3/2 |w|Hpw

w − w 1 (Γ ) + c2 h
3/2
wL2 (Γ )
is asymptotically of the same order of convergence as the corresponding error

estimate (14.3) of a standard Galerkin boundary element method. A com-
parison with the condition (14.17) which was needed to ensure the positive
definiteness of the approximate stiffness matrix Vh shows, that asymptotically
condition (14.17) follows from (14.18).
It remains to find suitable representations (14.11) of the fundamental so-
1
lution U ∗ (x, y) = − 2π log |x − y| satisfying the error estimate (14.12). Then,
the condition (14.18) also implies a suitable choice of the parameters η and .
14.2.1 Taylor Series Representations
A first possibility to derive a representation (14.11) is to consider a Taylor

expansion of the fundamental solution U ∗ (x, y) with respect to the integration
variable y ∈ ωiλ [73]. First we consider the Taylor expansion of a scalar function
f (t). For p ∈ N we have
p 1
1 dn 1 p d
p+1
f (1) = f (0) + f (t)|t=0 + (1 − s) f (s)ds.
n=1
n! dtn p! dsp+1
0
Let yiλ be the center of the cluster ωiλ . For an arbitrary y ∈ ωiλ and t ∈ [0, 1]
let
f (t) := U ∗ (x, yiλ + t(y − yiλ )).
Then we have
2
d ∂ ∗
f (t) = (yj − yi,j
λ
) U (x, z)|z=yiλ +t(y−yiλ ) ,
dt j=1
∂z j
and by applying this recursively, we obtain for 1 ≤ n ≤ p

dn n!
f (t) = (y − yiλ )α Dzα U ∗ (x, z)|z=yiλ +t(y−yiλ ) .
dtn α!
|α|=n
Thus, the Taylor expansion of the fundamental solution U ∗ (x, y) with respect
to the cluster center yiλ gives the representation
U ∗ (x, y) = U∗ (x, y) + Rp (x, y)
where
p
1
U∗ (x, y) = U ∗ (x, yiλ ) + (y − yiλ )α Dzα U ∗ (x, z)|z=yiλ (14.19)
α!
n=1 |α|=n
defines an approximation of the fundamental solution. By setting
f0λ,j (x) := U ∗ (x, yiλ ), g0λ,i (y) := 1
and
1
λ,j
fn,α (x) := Dzα U ∗ (x, z)|z=yiλ , (y − yiλ )α
λ,i
gn,α (y) :=
α!
for n = 1, . . . , p and |α| = n, we then obtain the representation (14.11). For
any n ∈ [1, p] there exist n + 1 multi–indices α ∈ N20 satisfying |α| = n. Then,
the number of terms in the series representation (14.11) is

p
1
= 1+ (n + 1) = (p + 1)(p + 2).
n=1
2
To derive the error estimate (14.12) we have to consider the remainder
1
1
Rp (y, yiλ ) = (1 − s)p (y − yiλ )α Dzα U ∗ (x, z)|z=yiλ +s(y−yiλ ) ds.
p!
0 |α|=p+1
For this we first need to estimate certain derivatives of the fundamental solu-
tion U ∗ (x, y).
Lemma 14.4. Let ωiλ and ωjλ be a pair of maximally admissible clusters. For
|α| = p ∈ N there holds the estimate
α ∗
Dy U (x, y) ≤ 1 3 (p − 1)!
p−1
for all (x, y) ∈ ωjλ × ωiλ .
2π |x − y|p
Proof. For the derivatives of the function f (x, y) = log |x − y| we first have
∂ yi − xi
f (x, y) = for i = 1, 2.
∂yi |x − y|2
For the second order derivatives we further obtain

∂2 1 (yi − xi )2
2 f (x, y) = 2
−2 for i = 1, 2,
∂yi |x − y| |x − y|4
and
∂2 (y1 − x1 )(y2 − x2 )
f (x, y) = −2 .
∂y1 ∂y2 |x − y|4
In general we find for |α| = ∈ N a representation of the form
(y − x)β
Dyα f (x, y) = aβ (14.20)
|x − y||β|+
|β|≤
where aβ are some coefficients to be characterized. Hence we conclude

α 1 c
Dy f (x, y) ≤ aβ = .
|x − y| |x − y|
|β|≤
A comparison with the first and second order derivatives of f (x, y) gives c1 = 1
and c2 = 3. Now, a general estimate of the constant c for ≥ 2 follows by
induction. From (14.20) we obtain for i = 1, 2 and j = i
∂ α ∂ (y − x)β
Dy f (x, y) = aβ
∂yi ∂yi |x − y||β|+
|β|≤
(yi − xi )βi −1 (yj − xj )βj
= aβ βi
|x − y||β|+
|β|≤

(yi − xi )βi +1 (yj − xj )βj
−(|β| + ) .
|x − y||β|++2
By using
|yi − xi | ≤ |x − y|, βi ≤ |β| ≤ , βi + βj ≤ |β| for i = j
we then obtain

∂ α 3 3c c+1

∂yi Dy f (x, y) ≤ |x − y|+1 aβ =
|x − y|+1
=
|x − y|+1
|β|≤
and therefore
c+1 = 3c = 3 ! .
In particular for = p this gives the assertion.

By using Lemma 14.4 we now can derive an estimate of the remainder
Rp (x, y) of the Taylor series approximation (14.19).
Lemma 14.5. Let ωiλ and ωjλ be a pair of maximally admissible clusters. For
the approximation U∗ (x, y) of the fundamental solution U ∗ (x, y) as defined in
(14.19) there holds the error estimate
p+1
∗
U (x, y) − U∗ (x, y) ≤ 3p 1 for all (x, y) ∈ ωjλ × ωiλ .
η
Proof. By applying Lemma 14.4 for (x, y) ∈ ωjλ × ωiλ and by using the admis-
sibility condition (14.8) we obtain
∗
U (x, y) − U∗ (x, y) = Rp (y, yiλ )
1
1
≤ |y − yiλ |p+1 max |Dzα U ∗ (x, z)z=ȳ | (1 − s)p ds
p! ȳ∈ωiλ
|α|=p+1 0
1 3p p!
= |y − yiλ |p+1 max
(p + 1)! ȳ∈ωiλ |x − ȳ|p+1
|α|=p+1
p+1
[diam ωiλ ]p+1 1
≤ 3p ≤ 3p .

[dist(ωiλ , ωjλ )]p+1 η
The decomposition (14.19) of the fundamental solution defines via (14.15) an

approximated stiffness matrix Vh . To ensure the asymptotically optimal error
estimate (14.3) the related condition (14.18) reads
p+1
1
3p ≤ c h2 .
η
If we choose a fixed admissibility parameter η > 3, then due to h = O(1/N )
we therefore obtain the estimate
p+1
1 3 1
≤ c 2
3 η N
from which we finally conclude
p = O(log2 N ) .
The total amount of work (14.16) to store Vh and to realize a matrix by vector
product with the approximated stiffness matrix Vh is then proportional to
N (log2 N )3 .
14.2.2 Series Representations of the Fundamental Solution

Instead of a Taylor expansion of the fundamental solution U ∗ (x, y) one may
derive alternative series expansions which are valid in the far field for (x, y) ∈
ωjλ × ωiλ where ωiλ and ωjλ is a pair of admissible clusters. In particular, these
series expansions are the starting point to derive the Fast Multipole Method
[64, 65] which combines the series expansions on different levels. However, the
resulting hierarchical algorithms will not discussed here, see, e.g. [48, 62, 109,
110].
Let yiλ be the center of the cluster ωiλ . By considering the fundamental
solution in the complex plane we have

1 1 1
− log |x − y| = − log |(x − yi ) − (y − yi )| = Re −
λ λ
log(z − z0 )
2π 2π 2π
where
) = |x − yiλ | eiϕ(x−yi ) ,
λ
z := (x1 − yi,1
λ
) + i(x2 − yi,2
λ
) = |y − yiλ | eiϕ(y−yi ) .
λ
z0 := (y1 − yi,1
λ
) + i(y2 − yi,2
λ
Since the clusters ωiλ and ωjλ are assumed to be admissible, it follows that
|z0 | |y − yiλ | diam ωiλ 1
= ≤ ≤ < 1.
|z| |x − yi |
λ dist(ωiλ , ωjλ ) η
Hence we can apply the series representation of the logarithm,
1 1 1 9 z0 :
− log(z − z0 ) = − log z − log 1 −
2π 2π 2π z
1 z 0 :n
9
∞
1 1
=− log z + ,
2π 2π n=1 n z
and for p ∈ N we can define an approximation
& '
1 1 9 z 0 :n
p
∗ 1
U (x, y) := Re − log z + (14.21)
2π 2π n=1 n z
of the fundamental solution U ∗ (x, y). By using
9 z :n |y − yiλ |n inϕ(y−yiλ ) −inϕ(x−yiλ )
0
= z0n z −n = e e
z |x − yiλ |n
we then obtain the representation
1
U∗ (x, y) = − log |x − yiλ |
2π
1 1
p
cos nϕ(x − yiλ )
+ |y − yiλ |n cos nϕ(y − yiλ )
2π n=1 n |x − yiλ |n
1 1
p
sin nϕ(x − yiλ )
+ |y − yiλ |n sin nϕ(y − yiλ ) .
2π n=1 n |x − yiλ |n
By introducing
f0λ,j (x) := U ∗ (x, yiλ ), g0λ,i (y) := 1
and
λ,j 1 cos nϕ(x − yiλ ) λ,i 1

f2n−1 (x) := , g2n−1 (y) := |y − yiλ |n cos nϕ(y − yiλ ),
2π |x − yiλ |n n
as well as
λ,j 1 sin nϕ(x − yiλ ) λ,i 1

f2n (x) := , g2n (y) := |y − yiλ |n sin nϕ(y − yiλ )
2π |x − yiλ |n n
for n = 1, . . . , p we finally obtain the representation (14.11) where = 2p + 1.
Lemma 14.6. Let ωiλ and ωjλ be a pair of admissible clusters. For the approx-
imation U (x, y) of the fundamental solution U ∗ (x, y) as defined in (14.21)
p
∗ ∗ 1 1 1 1
|U (x, y) − U (x, y)| ≤ for all (x, y) ∈ ωjλ × ωiλ .
2π p + 1 η − 1 η
Proof. By using the series expansion
1 1 9 z0 : n
∞
1 1
− log(z − z0 ) = − log z +
2π 2π 2π n=1 n z
we conclude from the admissibility condition (14.8)

& '
1 1 9 z0 :n
∞
∗ ∗
|U (x, y) − U (x, y)| = Re
2π n=p+1 n z
∞ n
1 1 1
≤
2π n=p+1 n η
p+1 ∞ n
1 1 1 1
≤ .

2π p + 1 η n=0
η
To ensure the asymptotically optimal error estimate (14.3) the related condi-
tion (14.18) now reads
p
1 1 1 1
≤ c h2 .
2π p + 1 η − 1 η
If we choose a fixed admissibility parameter η > 1 this finally gives
p = O(log2 N ) .
The total amount of work (14.16) to store Vh and to realize a matrix by vector
multiplication with the approximated stiffness matrix Vh is then proportional
to
N (log2 N )2 .
Note that both the approximation (14.19) based on the Taylor expansion
as well as the series expansion (14.21) define nonsymmetric approximations
U∗ (x, y) of the fundamental solution U ∗ (x, y), and therefore this results in
a nonsymmetric approximated stiffness matrix Vh . Hence we aim to derive a
symmetric approximation U∗ (x, y). For this we first consider the representa-
tion (14.21),
& '
1 1 p
1 9 z :n
0
U∗ (x, y) = Re − log z +
2π 2π n=1 n z
where
z = |x − yiλ |eiϕ(x−yi ) , z0 = |y − yiλ |eiϕ(y−yi ) .
λ λ
For the center yjλ of the cluster ωjλ we consider z = w − z1 where
w := |yjλ − yiλ |eiϕ(yj −yi ) , z1 := |yjλ − x|eiϕ(yj −x) .

λ λ λ
By using the admissibility condition (14.8) we have
|z1 | |yjλ − x| diam ωjλ 1

= λ ≤ ≤ < 1,
|w| |yj − yi |
λ λ λ
dist(ωi , ωj ) η
and therefore we can write
1 1 9 z1 : n
∞
1 1
− log z = − log w + .
2π 2π 2π n=1 n w
Lemma 14.7. Let w, z1 ∈ C satisfying |z1 | < |w|. For n ∈ N then there holds
∞
1 m+n−1 z1m
= .
(w − z1 )n n−1 wm+n
m=0
Proof. For |z1 | < |w| we first have
1 9 z1 :m
∞
1 1 1
= =
w − z1 w 1 − z1 w m=0 w
w
and therefore the assertion in the case n = 1. For n > 1 we have
1 1 dn−1 1
=
(w − z1 )n (n − 1)! dz1n−1 w − z1
# $
dn−1 1 9 z1 :m
∞
1
=
(n − 1)! dz1n−1 w m=0 w
∞

1 1 m! z1m−n+1
=
(n − 1)! w m=n−1
(m − n + 1)! wm
∞
1 1 (m + n − 1)! z1m
=
(n − 1)! w m=0 m! wm+n−1
∞
m+n−1 z1m
= .
n−1 wm+n
m=0
By using (14.21), z = w − z1 , and by applying Lemma 14.7 for p ∈ N we can

define a symmetric approximation of the fundamental solution U ∗ (x, y) as
&
1 1 9 z1 :m 1 1 9 z 0 :n
p p
∗ 1
U (x, y) = Re − log w + + (14.22)
2π 2π m=1 m w 2π n=1 n w
'
1 1 m + n − 1 z0n z1m
p p
+ .
2π n=1 m=1 n n−1 wm+n
Lemma 14.8. Let ωiλ and ωjλ be a pair of admissible clusters. For the approx-
imation U ∗ (x, y) of the fundamental solution U ∗ (x, y) as defined in (14.22)

p
∗ ∗ (x, y) ≤ 1 1 1 1 1
U (x, y) − U 2 +

2π p + 1 η − 1 η−1 η
for all (x, y) ∈ ωjλ × ωiλ .
Proof. By applying the triangle inequality and Lemma 14.6 we have

∗ ∗ (x, y) ≤ U ∗ (x, y) − U ∗ (x, y) + U ∗ (x, y) − U
∗ (x, y)
U (x, y) − U
p+1
1 1 η 1
≤
2π p + 1 η − 1 η
& '
1 1 9 z 1 :m
∞
1 1 m − n − 1 z0n z1m
p ∞

+ Re +
2π m=p+1 m w 2π n=1 m=p+1 n n−1 wm+n
p+1 ∞ m+n
1 1 m−n−1
p
1 1 η 1 1
≤ +
πp+1η−1 η 2π n=1 m=p+1 n n−1 η
where we have used the admissibility condition (14.8). With


1 m−n−1 1 (m − n − 1)! 1 (m − n − 1)! 1
= = ≤
n n−1 n (n − 1)!m! m n!(m − 1)! m
we obtain for the remaining term
∞ m+n ∞ m+n
1 1 m−n−1 1 1 1
p p
1
≤
2π n=1 m=p+1 n n−1 η 2π n=1 m=p+1 m η
p+1
1 1 η 1
≤
2π p + 1 (η − 1)2 η
which concludes the proof.

14.2.3 Adaptive Cross Approximation
All approximations U∗ (x, y) of the fundamental solution U ∗ (x, y) as described

before require either the knowledge of a suitable series expansion, or the com-
putation of higher order derivatives of the fundamental solution. Hence we
want to define approximations which only require the evaluation of the fun-
damental solution in appropriate interpolation nodes. One possibility is to
consider the Tschebyscheff interpolation of the fundamental solution. Here
we describe an alternative interpolation algorithm which was first given by
Tyrtyshnikov in [151], see also [12, 13].
Let ωiλ and ωjλ be a pair of admissible clusters. To define an approximation
of the fundamental solution U ∗ (x, y) for arguments (x, y) ∈ ωjλ × ωiλ we con-
sider two sequences of functions sk (x, y) and rk (x, y). In particular, rk (x, y)
describes the residual of the associated approximation sk (x, y). To initialize
this construction we first define
s0 (x, y) := 0, r0 (x, y) := U ∗ (x, y).
For k = 1, 2, . . . , let (xk , yk ) ∈ ωjλ × ωiλ be a pair of interpolation nodes with
a nonzero residual, i.e. αk := rk−1 (xk , yk ) = 0. Then we define the recursion
as
1
sk (x, y) := sk−1 (x, y) + rk−1 (x, yk )rk−1 (xk , y), (14.23)
αk
1
rk (x, y) := rk−1 (x, y) − rk−1 (x, yk )rk−1 (xk , y). (14.24)
αk
For ∈ N0 and for (x, y) ∈ ωjλ × ωiλ we finally define the approximation

rk−1 (x, yk )rk−1 (xk , y)
U∗ (x, y) = s (x, y) = (14.25)
rk−1 (xk , yk )
k=1
of the fundamental solution U ∗ (x, y) with respect to an admissible pair of

clusters (ωjλ , ωiλ ). The recursion as defined in (14.23) and (14.24) admits the
following properties.
Lemma 14.9. For 0 ≤ k ≤ and for all (x, y) ∈ ωjλ × ωiλ there holds
U ∗ (x, y) = sk (x, y) + rk (x, y). (14.26)
Moreover, there holds the interpolation property
rk (x, yi ) = 0 for all 1 ≤ i ≤ k (14.27)
as well as
rk (xj , y) = 0 for all 1 ≤ j ≤ k. (14.28)
Proof. By taking the sum of the recursions (14.23) and (14.24) we first have
rk (x, y) + sk (x, y) = rk−1 (x, y) + sk−1 (x, y)
for all k = 1, . . . , p, and therefore
rk (x, y) + sk (x, y) = r0 (x, y) + s0 (x, y) = U ∗ (x, y).
The interpolation properties follow by induction with respect to j, k. For k = 1

we have
1
r1 (x, y1 ) = r0 (x, y1 ) − r0 (x, y1 )r0 (x1 , y1 ) = 0.
r0 (x1 , y1 )
Hence we have rk (x, yi ) = 0 for k = 1, 2, . . . , p and i = 1, . . . , k. Then we
conclude
1
rk+1 (x, yi ) = rk (x, yi ) − rk (x, yk+1 )rk (xk+1 , yi ) = 0,
αk+1
i.e. rk+1 (x, yi ) = 0 for all i = 1, . . . , k. By using (14.24) we finally obtain
1
rk+1 (x, yk+1 ) = rk (x, yk+1 ) − rk (x, yk+1 )rk (xk+1 , yk+1 ) = 0.
rk (xk+1 , yk+1 )
The other interpolation property follows in the same way.
When inserting in (14.27) x = xj for j = 1, . . . , k this gives
rk (xj , yi ) = 0 for all i, j = 1, . . . , k,
and due to (14.26) we conclude
sk (xj , yi ) = U ∗ (xj , yi ) for all i, j = 1, . . . , k; k = 1, . . . , .
This means that the approximations sk (x, y) interpolate the fundamental so-
lution U ∗ (x, y) at the interpolation nodes (xj , yi ) for i, j = 1, . . . , k.
To analyze the approximations U∗ (x, y) as defined via (14.23) and (14.24)
we consider a sequence of matrices
Mk [j, i] = U ∗ (xj , yi ) for i, j = 1, . . . , k; k = 1, . . . , (14.29)
where we compute the determinants as follows.

Lemma 14.10. Let Mk , k = 1, . . . , , be the sequence of matrices as defined

in (14.29). Then there holds
,
k ,
k
det Mk = r0 (x1 , y1 ) · · · rk−1 (xk , yk ) = ri−1 (xi , yi ) = αi . (14.30)
i=1 i=1
Proof. To prove (14.30) we consider an induction with respect to k = 1, . . . , .

For k = 1 we first have
det M1 = M1 [1, 1] = U ∗ (x1 , y1 ) = r0 (x1 , y1 ) = α1 .
Assume that (14.30) holds for Mk . The matrix Mk+1 allows the representation
⎛ ∗ ⎞
U (x1 , y1 ) · · · U ∗ (x1 , yk ) U ∗ (x1 , yk+1 )
⎜ .. .. .. ⎟
⎜ . . . ⎟
Mk+1 = ⎜ ⎟.
⎝ U ∗ (xk , y1 ) · · · U ∗ (xk , yk ) U ∗ (xk , yk+1 ) ⎠
U ∗ (xk+1 , y1 ) · · · U ∗ (xk+1 , yk ) U ∗ (xk+1 , yk+1 )
For any (x, yi ) ∈ ωjλ × ωiλ we have by the recursion (14.24)
r0 (x, yi ) = U ∗ (x, yi ),
r0 (x, y1 )r0 (x1 , yi )
r1 (x, yi ) = r0 (x, yi ) −
r0 (x1 , y1 )
r0 (x1 , yi ) ∗
= U ∗ (x, yi ) − U (x, y1 )
r0 (x1 , y1 )
and therefore
r0 (x1 , yi )
r1 (x, yi ) = U ∗ (x, yi ) − α11 (yi ) U ∗ (x, y1 ), α11 (yi ) := .
r0 (x1 , y1 )
By induction, this representation can be generalized to all residuals rk (x, yi )

for all k = 1, . . . , and for all (x, yi ) ∈ ωjλ × ωiλ . Note that all residuals satisfy

k
rk (x, yi ) = U ∗ (x, yi ) − αjk (y)U ∗ (x, yj ) (14.31)
j=1
By using the recursion (14.24) and inserting twice the assumption of the
induction this gives
rk (x, yk+1 )rk (xk+1 , yi )

rk+1 (x, yi ) = rk (x, yi ) −
rk (xk+1 , yk+1 )

k
= U ∗ (x, yi ) − αjk (y)U ∗ (x, yj )
j=1
⎡ ⎤
rk (xk+1 , yi ) ⎣ ∗ k
− U (x, yk+1 ) − αjk (y)U ∗ (x, yj )⎦
rk (xk+1 , yk+1 ) j=1

k+1
∗
= U (x, yi ) − αjk+1 (y)U ∗ (x, yj ) .
j=1
By
-k+1 [j, i] := Mk+1 [j, i]
M for i = 1, . . . , k, j = 1, . . . , k + 1
and

k
-k+1 [j, k + 1] = Mk+1 [j, k + 1] −
M α k (y)Mk+1 [j, ] for j = 1, . . . , k + 1
=1
we define a transformed matrix M -k+1 satisfying det M -k+1 = det Mk+1 . Insert-
ing the definition of Mk+1 [j, ·] this gives for j = 1, . . . , k + 1, due to (14.31),

k
-k+1 [j, k + 1] = U ∗ (xj , yk+1 ) −
M α k (y)U ∗ (xj , y ) = rk (xj , yk+1 ).
=1
Note that, see Lemma 14.9,
rk (xj , yk+1 ) = 0 for all j = 1, . . . , k,
In particular, the determinant of the matrix Mk+1 remains unchanged when a

row Mk [j, ·] multiplied by αjk is subtracted from the last row of Mk . By using
(14.31) we then conclude
⎛ ∗ ⎞
U (x1 , y1 ) · · · U ∗ (x1 , yk ) 0
⎜ .. .. .. ⎟
-k+1 = ⎜
M ⎜ . . . ⎟
⎟.
⎝ U ∗ (xk , y1 ) · · · U ∗ (xk , yk ) 0 ⎠
U ∗ (xk+1 , y1 ) · · · U ∗ (xk+1 , yk ) rk (xk+1 , yk+1 )
-k+1 via an expansion with respect to the last column

The computation of det M
-k+1 now gives
of M
-k+1 = rk (xk+1 , yk+1 ) det Mk
det Mk+1 = det M
which concludes the induction.

By using the matrices Mk now we can represent the approximations
sk (x, y) as follows.
Lemma 14.11. For k = 1, . . . , the approximations sk (x, y) allow the repre-

sentations

sk (x, y) = (U ∗ (x, yi ))i=1,...,k Mk−1 (U ∗ (xi , y))i=1,...,k .
Proof. By using the recursion (14.23) the approximation sk (x, y) is defined as
k
1
sk (x, y) = ri−1 (x, yi )ri−1 (xi , y) .
α
i=1 i
From (14.31) we find for the residual ri−1 (x, yi ) the representation

i−1
∗
ri−1 (x, yi ) = U (x, yi ) − α i−1 (y)U ∗ (x, y ).
=1
Analogously we also have

i−1
ri−1 (xi , y) = U ∗ (xi , y) − β i−1 (x)U ∗ (x , y).
=1
Hence there exists a matrix Ak ∈ Rk×k satisfying
sk (x, y) = (U ∗ (x, y )) ∗

=1,...,k Ak U (x , y) =1,...,k .
In particular for (x, y) = (xj , yi ) we have rk (xj , yi ) = 0 and therefore
U ∗ (xj , yi ) = sk (xj , yi ) = (U ∗ (xj , y )) ∗

=1,...,k Ak U (x , yi ) =1,...,k
for i, j = 1, . . . , k. This is equivalent to
Mk = Mk Ak Mk
and since Mk is invertible this gives Ak = Mk−1 .

To estimate the residual r (x, y) of the approximation U∗ (x, y) as defined
in (14.25) we will consider a relation of the above approach with an inter-
polation by using Lagrange polynomials. For p ∈ N0 let Pp (R2 ) denote the
space of polynomials y α of degree |α| ≤ p where y ∈ R2 . We assume that the
number of interpolation nodes (xk , yk ) to define (14.25) corresponds with the
dimension of Pp (R2 ), i.e.
1
:= dim Pp (R2 ) = p(p + 1).
2
For k = let yk = y , then the Lagrange polynomials Lk ∈ Pp (R2 ) are well
defined for k = 1, . . . , , and we have
Lk (y ) = δk for k, = 1, . . . , .
The Lagrange interpolation of the fundamental solution U ∗ (x, y) is then given

as

∗
UL, (x, y) = U ∗ (x, yk )Lk (y) = (U ∗ (x, yk ))k=1,..., (Lk (y))k=1,...,
k=1
where the associated residual is

E (x, y) := U ∗ (x, y) − (U ∗ (x, yk ))k=1,..., (Lk (y))k=1,..., .
Let ⎛ ⎞
U ∗ (x1 , y1 ) · · · · · · U ∗ (x1 , y )
⎜ .. .. ⎟
⎜. . ⎟
⎜ ∗ ⎟
⎜ U (x −1 , y1 ) ······ U (x −1 , y ) ⎟
∗
⎜ ∗ ⎟
M, (x) := ⎜
⎜ U ∗ (x, y1 ) ······ U ∗ (x, y ) ⎟.
⎟
⎜ U (x +1 , y1 ) ······ U (x +1 , y ) ⎟
∗
⎜ ⎟
⎜. .. ⎟
⎝ .. . ⎠
U ∗ (x , y1 ) ······ ∗
U (x , y )
Lemma 14.12. For (x, y) ∈ ωjλ × ωiκ let r (x, y) be the residual of the ap-
proximation U∗ (x, y). Then there holds

det M,k (x)
r (x, y) = E (x, y) − E (xk , y).
det M
k=1
Proof. By using Lemma 14.9 and the matrix representation of s (x, y), see
Lemma 14.11, we first have
r (x, y) = U ∗ (x, y) − s (x, y)

= U ∗ (x, y) − (U ∗ (x, yk ))k=1, M−1 (U ∗ (xk , y))k=1,

= U ∗ (x, y) − (U ∗ (x, yk ))k=1, (Lk (y))k=1,
2 3

− (U ∗ (x, yk ))k=1, M−1 (U ∗ (xk , y))k=1, − M (Lk (y))k=1, .
Due to
& '

∗ ∗ ∗
(U (xk , y))k=1, − M (Lk (y))k=1, = U (xk , y) − U (xk , y )L (y)
=1 k=1,
= (E (xk , y))k=1,
we then conclude

r (x, y) = E (x, y) − (U ∗ (x, yk ))k=1, M−1 (E (xk , y))k=1,
9 :
= E (x, y) − M− (U ∗ (x, y )) =1, (E (xk , y))k=1, .
Let e ∈ R be the unit vectors satisfying e i = δi . Then,
M ei = (U ∗ (xi , yk ))k=1,
and thus
ei = M− (U ∗ (xi , yk ))k=1, .
Hence we conclude
M− M,
(x) =
9 :
= M− (U ∗ (xi , yk ))k=1,;i=1, −1 , (U ∗ (x, yk ))k=1, , (U ∗ (xi , yk ))k=1,;i= +1,
9 :
= e1 , . . . , e −1 , M− (U ∗ (x, yk ))k=1, , e +1 , . . . , e ,
and with
9 : det M,k (x)
M− (U ∗ (x, y ) =1, = det M− M,k

(x) =
k det M
we finally get the assertion.

Corollary 14.13. Let the interpolation nodes (xk , yk ) ∈ ωjλ × ωiλ be chosen
such that
|det M, (x)| ≤ |det M | (14.32)
is satisfied for all = 1, . . . , and for all x ∈ ωjλ . The residual r (x, y) then
satisfies the estimate
|r (x, y)| ≤ (1 + ) sup |E (x, y)| .

x∈ωjλ
The criteria (14.32) to define the interpolation nodes (xk , yk ) ∈ ωjλ ∈ ωiλ seems
not be very suitable for a practical realization. Hence we finally consider an
alternative choice.
Lemma 14.14. For k = 1, . . . , let the nodal pairs (xk , yk ) ∈ ωjλ ∈ ωiλ be
chosen such that
|rk−1 (xk , yk )| ≥ |rk−1 (x, yk )| for all x ∈ ωjλ (14.33)
is satisfied. Then there holds

|det Mk, (x)|
sup ≤ 2k− .
x∈ωjλ |det Mk |
Proof. As in the proof of Lemma 14.10 we have for det Mk, (x) and 1 ≤ < k
the recursion
det Mk, (x) = rk−1 (xk , yk )det Mk−1, (x) − rk−1 (x, yk )det Mk−1, (xk ),
14.3 Wavelets 351
or,
det M1,1 (x) = r0 (x, y1 ), det Mk,k (x) = rk−1 (x, yk )det Mk−1
for k = 2, 3, . . . , , as well as
det Mk = rk−1 (xk , yk ) det Mk−1 .
Hence we have for 1 ≤ < k by using the assumption (14.33)
det Mk, (x) det Mk−1, (x) rk−1 (x, yk ) det Mk−1, (x)
= −
det Mk det Mk−1 rk−1 (xk , yk ) det Mk−1
and therefore
|det Mk, (x)| |det Mk−1, (x)|
sup ≤ 2 sup ,
x∈ωjλ |det Mk | x∈ωjλ |det Mk−1 |
or,
det Mk,k (x)
= rk−1 (x, yk ) ≤ 1 .

det Mk rk−1 (xk , yk )
By using Lemma 14.14 we obtain from Lemma 14.12 an upper bound of

the residual r (x, y) by the Lagrange interpolation error E (x, y).
Corollary 14.15. For k = 1, . . . , let the nodal pairs (xk , yk ) ∈ ωjλ × ωiλ be
chosen such that assumption (14.33) is satisfied. Then there holds
|r (x, y)| ≤ 2 sup |E (x, y)| .

x∈ωjλ
Contrary to the one–dimensional Lagrange interpolation the interpolation

in more space dimensions is quite difficult. In particular the uniqueness of the
interpolation polynomial depends on the choice of the interpolation nodes.
Moreover, there is no explicit representation of the remainder E (x, y) known.
Hence we skip a more detailed discussion at this point. By using results of
[121] one can derive similar error estimates as in Lemma 14.6, see [13].
The adaptive cross approximation algorithm to approximate a scalar func-
tion as described in this subsection can be generalized in a straightforward
way to define low rank approximations of a matrix, see, e.g. [12, 14], and [117]
for a more detailed discussion.
14.3 Wavelets
In this subsection we introduce wavelets as hierarchical basis functions to
be used in the Galerkin discretization of the single layer potential V . As in
standard boundary element methods this leads to a dense stiffness matrix,
but by neglecting small matrix entries one can define a sparse approximation.
This reduces both the amount of storage, and amount to realize a matrix by
vector multiplication.
Without loss of generality we assume that the Lipschitz boundary Γ = ∂Ω
of a bounded domain Ω ⊂ R2 is given via a parametrization Γ = χ(Q) with
respect to a parameter domain Q = [0, 1] where we assume that the Jacobian
|χ̇(ξ)| is constant for all ξ ∈ Q. Moreover, we extend the parametrization
Γ = χ([0, 1]) periodically onto R. Hence we can assume the estimates
cχ1 |ξ − η| ≤ |χ(ξ) − χ(η)| ≤ cχ2 |ξ − η| for all ξ, η ∈ R (14.34)
with some positive constants cχ1 and cχ2 . In the case of a piecewise smooth
Lipschitz boundary Γ all following considerations have to be transfered to the
non–periodic parametrizations describing the parts Γj satisfying |χ̇j | = cj .
For j ∈ N we consider a decomposition of the parameter domain Q = [0, 1]
into Nj = 2j finite elements q j of mesh size |q j | = 2−j , = 1, . . . , Nj ,
2
*
j
Qj = q j , q j := (( − 1)2−j , 2−j ) for = 1, . . . , 2j .

=1
A decomposition Qj implies an associated trial space of piecewise constant

functions,
Vj := Sj0 (Q) = span{ϕ

N
j } =1
j
⊂ L2 (Q), dim Vj = Nj = 2j ,
where the basis functions are given as

j 1 for ξ ∈ q j ,
(x) =
ϕ
0 elsewhere.
By construction we have the nested inclusions
V0 ⊂ V1 ⊂ · · · ⊂ VL = SL0 (Q) ⊂ VL+1 ⊂ · · · ⊂ L2 (Q).
For any j > 0 we now construct subspaces Wj as L2 (Q)–orthogonal comple-

ments of Vj−1 in Vj , i.e.
! "
W0 := V0 , Wj := ϕ j ∈ Vj : ϕ
j , ϕ
j−1
i j−1
L2 (Q) = 0 for all ϕ i ∈ Vj−1
where
dimW0 = 1, dimWj = dimVj − dimVj−1 = 2j−1 for j > 0.
Hence we obtain a multilevel decomposition of the trial space VL = SL0 (Q) as
VL = W0 ⊕ W1 ⊕ · · · ⊕ WL .
Due to W0 = V0 we have
14.3 Wavelets 353
W0 = span{ψ10 } where ψ10 (ξ) = 1 for ξ ∈ Q = [0, 1], ψ10 L2 (Q) = 1.
It remains to construct a basis of
Wj = span{ψ j }2 =1
j−1
for j = 1, . . . , L.
By using dim W1 = 1 we have to determine one basis function only. By setting
ψ11 (ξ) = a1 ϕ
11 (ξ) + a2 ϕ
12 (ξ) for ξ ∈ Q = [0, 1]
we obtain from the orthogonality condition

1
0 = ψ10 (ξ)ψ11 (ξ)dξ = (a1 + a2 )
2
Q
and therefore a2 = −a1 . Hence we can define the basis function as

1 for ξ ∈ q11 = (0, 12 ),
ψ1 (ξ) =
1
ψ11 L2 (Q) = 1.
−1 for ξ ∈ q21 = ( 12 , 1),
By applying this recursively we obtain the following representation of the basis

functions,
⎧
⎪
⎪ 1 for ξ ∈ ((2 − 2)2−j , (2 − 1)2−j ),
⎨
ψ% j (ξ) = −1 for ξ ∈ ((2 − 1)2−j , 2 2−j ),
⎪
⎪
⎩
0 elsewhere
where = 1, . . . , 2j−1 , j = 2, 3, . . .. Moreover, we have

ψ% j 2L2 (Q) = supp ψ% j = 21−j for j ≥ 1.
Hence we can define normalized basis functions as
ψ j (ξ) = 2(j−1)/2 ψ% j (ξ) for = 1, . . . , 2j−1 , j ≥ 1. (14.35)
Due to the orthogonality relation

ψ j (ξ)ψ10 (ξ)dξ = 0 for all = 1, . . . , 2j−1 , j ≥ 1
Q
we also obtain the moment condition

ψ j (ξ)dξ = 0 for all = 1, . . . , 2j−1 , j ≥ 1 (14.36)
Q
which holds for piecewise constant wavelets ψ j . The basis functions (14.35)
as constructed above are also denoted as Haar wavelets, see Fig. 14.4.
1 ψ10
-
0 1
−1 ψ11
ψ12 ψ22
Fig. 14.4. Piecewise constant wavelets for j = 0, 1, 2.
Via the parametrization Γ = χ(Q) we can also define boundary elements

τ j = χ(q j ) of the mesh size

hj = dsx = |χ̇(ξ)|dξ = |χ̇| 2−j .
τj qj
<Nj j
The global mesh size of the boundary element mesh ΓNj = =1 τ is then
given as hj = |χ̇| 2−j . Moreover, we can lift both the piecewise constant basis
j ∈ Vj as well as the wavelets ψ j ∈ Wj on the boundary Γ = ∂Ω,
functions ϕ
for x = χ(ξ) ∈ Γ we have
ϕj (x) := ϕ
j (ξ), ψ j (x) := ψ j (ξ) for ξ ∈ Q = [0, 1].
For j > 0 these basis functions define the trial spaces

max{1,2j−1 }
Vj = span{ϕj }2 =1 ,
j
Wj = span{ψ j } =1 ,
and for L ∈ N we have

14.3 Wavelets 355
VL = Sh0L (Γ ) = span{ϕL j
} =1 = span{ψ } =1,...,max{1,2j−1 },j=1,...,L ,
NL
i.e. any function whL ∈ VL can be described as
L max{1,2 }
i−1

whL = wki ψki ∈ VL . (14.37)
i=0 k=1
Due to

ψki , ψ j L2 (Ω) = ψki (x)ψ j (x)dsx
Γ

|χ̇| for i = j, k = ,
= |χ̇| ψki (ξ)ψ j (ξ)dξ =
0 elsewhere
Q
the orthogonality of the trial spaces Wki in the parameter domain is transfered
to the trial spaces Wki which are defined with respect to the boundary element
mesh.
By using Remark 13.21 we can derive spectrally equivalent norm represen-
tations by means of the multilevel representation (14.37) of a given function
whL ∈ VL .
Lemma 14.16. Let whL ∈ VL be given as in (14.37). Then,
max{1,2i−1 }

L i 2
whL 2L,s = 22si wk .
i=0 k=1
defines an equivalent norm in H s (Γ ) for all s ∈ (− 12 , 12 ).
Proof. Since whL ∈ VL = Sh0L (Γ ) is a piecewise constant function, and by

using Remark 13.21, the bilinear form of the multilevel operator B s ,

L
B whL , whL L2 (Γ ) =
s
h−2s
i (Qi − Qi−1 )whL 2L2 (Γ ) ,
i=0
defines an equivalent norm in H s (Γ ), s ∈ (− 12 , 12 ). Thereby, the L2 projection

Qi whL ∈ Vi = Sh0i (Γ ) is the unique solution of the variational problem
Qi whL , vhi L2 (Γ ) = whL , vhi L2 (Γ ) for all vhi ∈ Vi .
By using the orthogonality of basis functions we conclude for the L2 projection

the representation
max{1,2j−1 } i max{1,2
j−1
}
1
i
Qi whL = whL , ψ j L2 (Γ ) ψ j = w j ψ j .
|χ̇| j=0 j=0
=1 =1
When taking the difference of two succeeding L2 projections this gives
max{1,2i−1 }

(Qi − Qi−1 )whL = wki ψki
k=1
and therefore
= =2
=max{1,2i−1 } =

L
= =
−2s = i i=
B s whL , whL L2 (Γ ) = hi = wk ψk =
i=0 = k=1 =
L2 (Γ )
max{1,2i−1 }

L i 2
= |χ̇| h−2s wk .
i
i=0 k=1
By inserting the mesh sizes hi = |χ̇| 2−i this concludes the proof.
By using the representation (14.37) the variational problem (14.1) is equiv-
alent to
L max{1,2 }
i−1

wki V ψki , ψ j Γ = g, ψ j Γ for all ψ j ∈ VL . (14.38)
i=0 k=1
Hence we have to compute the entries of the stiffness matrix

1 j
V [(, j), (k, i)] = −
L
ψ (x) log |x − y|ψki (y)dsy dsx
2π
Γ Γ
for i = 1, . . . , max{1, 2 −1 }, j = 1, . . . , max{1, 2k−1 } and k, = 0, . . . , L. Now

we can estimate the matrix entries V L [(, j), (k, i)] when assuming a certain
relation of the supports of the basis functions ψki and ψ j . For this we first
define the support of ψki as

Ski := supp ψki ⊂ Γ,
and 9 :
dij i j
k := dist Sk , S = min |x − y|
(x,y)∈Ski ×Sj
describes the distance between the supports of the basis functions ψki and ψ j ,
respectively.
Lemma 14.17. Assume that for i, j ≥ 2 the condition dij k > 0 is satisfied.
Then there holds the estimate
|χ̇|4 −3(i+j)/2 9 ij :−2
|V L [(, j), (k, i)]| ≤ 2 dk .
2π
14.3 Wavelets 357
Proof. By inserting the parametrization Γ = χ(Q) and by using the nor-

malized basis functions (14.35) the entries of the stiffness matrix V L can be
computed as

1 j
V [(, j), (k, i)] = −
L
ψ (x) log |x − y|ψki (y)dsy dsx
2π
Γ Γ
1
= − ψ (ξ) log |χ(η) − χ(ξ)| ψki (η) |χ̇(η)| dη |χ̇(ξ)| dξ
j
2π
Q Q

|χ̇|2 (i−1)/2 (j−1)/2
= − 2 2 ψ% j (ξ) log |χ(η) − χ(ξ)| ψ%ki (η) dη dξ.
2π
Q Q
With the substitutions
η = η(s) = (2k − 2)2−i + s 21−i , ξ = ξ(t) = (2 − 2)2−j + t 21−j
for s, t ∈ Q = [0, 1] this is equivalent to

|χ̇|2 −(i+j)/2
V [(, j), (k, i)] = −
L
2 ψ%11 (ξ(t)) k(s, t)ψ%11 (η(s)) ds dt
π
Q Q
where the kernel function is given by
k(s, t) = log |χ(η(s)) − χ(ξ(t))| .
Due to the moment condition (14.36) we can replace the kernel function k(s, t)
by r(s, t) := k(s, t) − P1 (s) − P2 (t) where P1 (s) and P2 (t) correspond to
the first terms of the Taylor expansion of k(s, t), i.e. r(s, t) corresponds to
the remainder of the Taylor expansion. The Taylor expansion of the kernel
function k(s, t) with respect to s0 = 12 gives

1 1 ∂
k(s, t) = k ,t + s − k(s, t)
2 2 ∂s s=s̄
with a suitable s̄ ∈ Q. Applying another Taylor expansion with respect to

t0 = 12 we obtain
2
∂ ∂ 1 1 ∂
k(s, t) = k s, + t− k(s, t)
∂s s=s̄ ∂s 2 s=s̄ 2 ∂s∂t (s,t)=(s̄,t̄)
where t̄ ∈ Q. Hence we have

1 1 ∂ 1
k(s, t) − k ,t − s − k s,
2 2 ∂s 2 s=s̄
2
1 1 ∂
= s− t− k(s, t) .
2 2 ∂s∂t (s,t)=(s̄,t̄)
Due to the moment condition (14.36) we conclude

1 ∂ 1 1
ψ%11 (ξ(t)) s − k s, dt = ψ%11 (η(s))k , t ds = 0 .
2 ∂s 2 s=s̄ 2
Q Q
Hence we obtain

|χ̇|2 −(i+j)/2
V [(, j), (k, i)] = −
L
2 ψ%11 (ξ(t)) k(s, t)ψ%11 (η(s)) ds dt
π
Q Q
2
|χ̇| −(i+j)/2 1 1
=− 2 ψ%11 (ξ(t)) s− t− ·
π 2 2
Q Q

∂2
· k(s, t) ψ%1 (η(s)) ds dt
1
∂s∂t (s,t)=(s̄,t̄)
and therefore
2
L 2 ∂
V [(, j), (k, i)] ≤ |χ̇| 2−(i+j)/2 1
max k(s, t).
π 16
(s,t)∈Q×Q ∂s∂t
By applying the chain rule we further have

∂2 ∂2
k(s, t) = 21−i 21−j log |χ(η) − χ(ξ)| .
∂s∂t ∂η∂ξ
Moreover,
2
∂ ∂ ∂
log |χ(η) − χ(ξ)| = log |y − x(ξ)||y=χ(η) χj (η),
∂η i=1
∂y i ∂η
as well as
2
∂2 ∂2 ∂ ∂
log |χ(η)−χ(ξ)| = log |y−x||y=χ(η),x=χ(ξ) χi (η) χi (ξ) .
∂η∂ξ i,j=1
∂y i ∂xj ∂η ∂ξ
Applying the Cauchy–Schwarz inequality twice this gives

⎛ ⎞
2 2
2 1/2
∂ ∂ 2
2⎝
log |y − x||y=χ(η),x=χ(ξ) ⎠ .
∂η∂ξ log |χ(η) − χ(ξ)| ≤ |χ̇| ∂yi ∂xj
i,j=1
By using
∂2 1 (xi − yi )2
log |x − y| = − 2
+2 ,
∂xi ∂yi |x − y| |x − y|4
∂2 (x1 − y1 )(x2 − y2 )
log |x − y| = 2
∂x1 ∂y2 |x − y|4
14.3 Wavelets 359
we finally obtain
2
∂ 2
max k(s, t) ≤ 22−(i+j) |χ̇|2 max .

(s,t)∈Q×Q ∂s∂t (x,y)∈Ski ×Sj |x − y|2
The estimate of Lemma 14.17 describes the decay of the matrix entries
V L [(, j), (k, i)] when considering wavelets ψki and ψ j with supports Ski and S j
which are far to each other. By defining an appropriate compression parameter
we therefore can characterize matrix entries V L [(, j), (k, i)] which can be
neglected when computing the stiffness matrix V L . For real valued parameter
α, κ ≥ 1 we first define a symmetric parameter matrix by
τij := α 2κL−i−j .
This enables the definition of a symmetric approximation V L of the stiffness

matrix V L by

V L [(, j), (k, i)] if dij
k ≤ τij ,
L
V [(, j), (k, i)] := (14.39)
0 elsewhere.
For the following considerations, in particular to estimate the number of

nonzero elements of the matrix V L as well as for the related stability and
error analysis we define for fixed i, j ≥ 2 block matrices

VijL := V L [(, j), (k, i)] k=1,...,2i−1 , =1,...,2j−1 ,
and the corresponding approximation VijL , respectively.
Lemma 14.18. The number of nonzero elements of the approximated stiff-

ness matrix V L as defined in (14.39) is O(N κ (log2 N )2 ).
Proof. For i = 0, 1 and j = 0, 1 the number of nonzero elements is 4(N − 1).

For i, j ≥ 2 we estimate the number of nonzero elements of the approximate
block matrix VijL as follows.
By using the parametrization Γ = χ(Q) we can identify the basis functions
ψki and ψ j with basis functions ψki and ψ j which are defined in the parameter
domain Q. For the support of the basis functions ψki and ψ j we then obtain
Ski = ((2(k − 1)2−i , 2k 2−i ), S j = (2( − 1)2−j , 2 2−j ).
For an arbitrary but fixed = 1, . . . , 2j−1 we first determine all basis functions
ψki where the supports Ski and S j do not overlap, i.e.
2k 2−i ≤ 2( − 1)2−j , 2 2−j ≤ 2(k − 1)2−i .
Then it follows that

1 ≤ k ≤ ( − 1)2i−j , 1 + 2i−j ≤ k ≤ 2i−1 .

Next we choose from the above those basis functions ψik which in addition
satisfy the distance condition
dij
k ≤ τij = α 2
κL−i−j
.
Due to assumption (14.34) we have dij χ i j

k ≤ c2 dist(Sk , S ). By requesting
cχ2 dist(Ski , S j ) ≤ α 2κL−i−j

we can find an upper bound for the number of related basis functions. By
using
α
0 < dist(Ski , S j ) = 2(k − 1)2−i − 2 2−j ≤ χ 2κL−i−j
c2
j
for S η < ξ ∈ S we obtain the estimate
i
k
α κL−j
k ≤ 1 + 2i−j + 2
2cχ2
and therefore

α κL−j
1 + 2i−j
≤ k ≤ min 2 , 1 + 2
i−1 i−j
+ χ2 .
2c2
Hence we can estimate the number of related nonzero elements as
α κL−j
2 .
2cχ2
This results follows analogously in the case Ski ξ < η ∈ S j . Thus, for a fixed
= 1, . . . , 2j−1 there exist maximal
α κL−j
2
cχ2
nonzero elements. The number of nonzero elements of the approximate block
matrix VijL is therefore bounded by
α κL−j α
2j−1 χ 2 = χ 2κL .
c2 2c2
By taking the sum over all block matrices VijL where k, = 2, . . . , L we can es-
timate, by taking into account the special situation for i, j = 0, 1, the number
of nonzero elements of V L by
α
4(N − 1) + (L − 1)2 χ 2κL .
2c2
A similar estimate follows when considering basis functions with overlapping
supports. When inserting N = 2L or L = log2 N this concludes the proof.
To estimate the approximation error V L − V L of the stiffness matrix V L
we first consider the approximation errors VijL − VijL of the block matrices
VijL .
14.3 Wavelets 361
Lemma 14.19. For i, j = 2, . . . , L let VijL be the exact Galerkin stiffness ma-
trix of the single layer potential V with respect to the trial spaces Wi and Wj .
Let VijL be the approximation as defined in (14.39). Then there hold the error
estimates
VijL − VijL ∞ ≤ c1 2−(i+j)/2 2−j (τij )−1 ,
VijL − VijL 1 ≤ c2 2−(i+j)/2 2−i (τij )−1 .
Proof. By using Lemma 14.17 we first have
2
i−1

L
VijL − VijL ∞ = maxj−1 V [(, j), (k, i)] − V L [(, j), (k, i)]
=1,...,2
k=1
2
i−1
L
= maxj−1 V [(, j), (k, i)]
=1,...,2
k=1
dij
k >τij
2
i−1
−3(i+j)/2 −2
≤ c2 max (dij
k ) .
=1,...,2j−1
k=1
dij
k >τij
Since we assume (14.34) to be satisfied for the parametrization of the boundary

Γ we then conclude
2
i−1
VijL − VijL ∞ ≤
c2 −3(i+j)/2
max (dist(Ski , S j ))−2 .
=1,...,2j−1
k=1
cχ i j
1 dist(Sk ,S )>τij
For an arbitrary but fixed = 1, . . . , 2j−1 the sum can be further estimated
by
2
i−1
−2
(dist(Ski , S j ))−2 ≤ 2 2(k − 1)2−i − 22−j
k=1 k>1+ α
χ 2κL−j + 2i−j
2c1
cχ i j
1 dist(Sk ,S )>τij
1
= 22i−1 .
n2
n> α
χ 2κL−j
2c1
Let n1 ∈ N be the smallest number satisfying n1 ≥ α κL−j

2cχ
2 . Then we have
1
∞
∞ ∞
1 1 1 1 1 1 1 2 4cχ1 j−κL
= + ≤ + dx = + ≤ ≤ 2
n=n
n2 n21 n=n +1 n2 n21 x2 n21 n1 n1 α
1 1 x=n1
which gives immediately the first estimate. The second estimate follows in the
same way.
By applying Lemma 13.17 (Schur Lemma) we can now estimate the error
of the approximation V L .
Theorem 14.20. For wh , vh ∈ VL = Sh0L (Γ ) ↔ w, v ∈ RN there holds the

error estimate

((V L − V L )w, v) ≤ c γ(hL , σ1 , σ2 ) wh H σ1 (Γ ) vh H σ2 (Γ )
where
⎧ κ+σ1 +σ2
⎪
⎪ hL for σ1 , σ2 ∈ (− 12 , 0),
⎪
⎨ hκ | ln h |
L L for σ1 = σ2 = 0,
γ(hL , σ1 , σ2 ) :=
⎪
⎪ h κ
for σ1 , σ2 ∈ (0, 12 ),
⎪ L
⎩ κ+σ2
hL for σ1 ∈ (0, 12 ), σ2 ∈ (− 12 , 0).
Proof. First we note that for i = 0, 1 and j = 0, 1 there is no approximation

of the matrix entries of V L . Then, by using the Cauchy–Schwarz inequality
and Lemma 14.16 we obtain

((V L − V L )w, v)

L 2 2
i−1 j−1
2 3
L
= wki v j V L [(, j) − (k, i)] − V L [(, j), (k, i)]
i=2 j=2 k=1 =1

L 2 2
i−1 j−1
L
= 2σ1 i wki 2σ2 j v j 2−σ1 i−σ2 j ·
i=2 j=2 k=1 =1

2 3

· V L [(, j) − (k, i)] − V L [(, j), (k, i)]

⎛ ⎞1/2 ⎛ ⎞1/2
L 2i−1
L 2j−1
2
2
≤ A2 ⎝ 22σ1 i wki ⎠ ⎝ 22σ2 j v j ⎠
i=2 k=1 j=2 =1
≤ A2 wh L,σ1 vh L,σ2
≤ A2 wh H σ1 (Γ ) vh H σ2 (Γ )
where the matrix A is defined by

2 3
A[(, j), (k, i)] = 2−σ1 i−σ2 j V L [(, j) − (k, i)] − V L [(, j), (k, i)] .
By using (13.17) we now can estimate the spectral norm A2 for an arbitrary
s as
14.3 Wavelets 363
L 2

i−1
A22 ≤ sup |A[(, j), (k, i)]|2s(i−j) ·

j = 2, . . . , L i=2 k=1
=1,...,2j−1
L 2

j−1
· sup |A[(, j), (k, i)]|2s(j−i) .

i = 2, . . . , L j=2 =1
k=1,...,2i−1
By using Lemma 14.19 we can bound the first term by

L 2

i−1

L s(i−j) −σ1 i−σ2 j
V [(, j), (k, i)] − V [(, j), (k, i)] 2
L
A1 = sup 2
j = 2, . . . , L i=2 k=1
=1,...,2j−1

L 2
i−1

L
V [(, j), (k, i)]− V [(, j), (k, i)]
s(i−j) −σ1 i−σ2 j L
= sup 2 2 sup
j=2,...,L i=2 =1,...,2j−1 k=1

L
= sup 2s(i−j) 2−σ1 i−σ2 j VijL − VijL ∞
j=2,...,L i=2

L
≤c sup 2s(i−j) 2−σ1 i−σ2 j 2−(i+j)/2 2−j (τij )−1
j=2,...,L i=2

L
1 i+j−κL
=c sup 2s(i−j) 2−σ1 i−σ2 j 2−(i+j)/2 2−j 2
j=2,...,L i=2 α
1
L
1
c 2−κL
= sup 2j(−s− 2 −1+1−σ2 ) 2i(s− 2 +1−σ1 )
j=2,...,L i=2

L
c 2−κL
= sup 2−σ2 j 2−σ1 i
j=2,...,L i=2
where s = − 12 . Note that

⎧ −σ L
⎪
⎨2
1
for σ1 ∈ (− 12 , 0),

L
2−σ1 i ≤ c L for σ1 = 0,
⎪
⎩
i=2 1 for σ1 ∈ (0, 12 )
and
−σ2 j 2−σ2 L for σ2 ∈ (− 12 , 0),
sup 2 ≤ c
j=2,...,L 1 for σ2 ∈ [0, 12 ).
To estimate the second term we proceed in an analogous way, and inserting
hL = 2−L finally gives the announced error estimate.
By using Theorem 8.3 (Strang Lemma) and Theorem 14.20 we can now
derive the stability and error analysis of the approximated stiffness matrix V L .
For this we first need to establish the positive definiteness of the approximated
stiffness matrix V L .
Theorem 14.21. For κ > 1 and for a sufficient small global boundary ele-
ment mesh size hL the approximated stiffness matrix V L is positive definite,
i.e.
1
(V L w, w) ≥ cV1 whL 2H −1/2 (Γ )
2
is satisfied for all whL ∈ VL ↔ w ∈ RN .
Proof. For σ ∈ (− 12 , 0) we have by using Theorem 14.20, the H −1/2 (Γ )–

ellipticity of the single layer potential V , and by applying the inverse inequality
in VL
(V L w, w) = (V L w, w) + ((V L − V L )w, w)

≥ V whL , whL Γ − |((V L − V L )w, w)|
≥ cV1 whL 2H −1/2 (Γ ) − c hκ+2σ
L whL 2H σ (Γ )
≥ cV1 whL 2H −1/2 (Γ ) − c hκ+2σ

L cI h−1−2σ
L whL 2H −1/2 (Γ )
0 1
= cV1 − c hκ−1
L wh 2H −1/2 (Γ ) .
Now, if
chκ−1
L ≤ 12 cV1 is satisfied the assertion follows.

Instead of the linear system V L w = f which corresponds to the variational
problem (14.38) we now have to solve the perturbed linear system V L w = f
where w ∈ RN ↔ w hL ∈ VL defines the associated approximate solution.
1
Theorem 14.22. Let w ∈ Hpw (Γ ) be the unique solution of the boundary
integral equation V w = g. For the approximate solution w hL ∈ VL ↔ w ∈ RN
of the perturbed linear system V L w
= f there holds the error estimate
3/2 κ−1/2
w − w
hL H −1/2 (Γ ) ≤ c1 hL wHpw
1 (Γ ) + c2 h
L wH 1/2 (Γ ) .
Proof. The solutions w, w ∈ RN of the linear systems V L w = f and V L w

=f
satisfy the orthogonality relation
(V L w − V L w,
v) = 0 for all v ∈ RN .
By using the positive definiteness of the approximated stiffness matrix V L we

then obtain
1 V
hL 2H −1/2 (Γ ) ≤ (V L (w − w),
c whL − w w − w)
2 1
= ((V L − V L )w, w − w).

By applying Theorem 14.20 we conclude for σ1 ∈ (0, 12 ) and σ2 ∈ (− 12 , 0)

14.3 Wavelets 365
1 V
c whL − w
hL H −1/2 (Γ ) ≤ c hκ+σ 2
whL H σ1 (Γ ) whL − w
hL H σ2 (Γ ) .
2 1 L
Further we have as in Lemma 12.2
whL H σ1 (Γ ) ≤ wH σ1 (Γ ) + w − whL H σ1 (Γ ) ≤ c wH σ1 (Γ ) ≤ c wH 1/2 (Γ ) .
On the other hand, by using the inverse inequality this gives

− 1 −σ2
whL − w
hL H σ2 (Γ ) ≤ cI hL 2 whL − w
hL H −1/2 (Γ ) .
Hence we have
1 V κ−1/2
c whL − w
hL H −1/2 (Γ ) ≤ c hL wH 1/2 (Γ ) .
2 1
Now the assertion follows from applying the triangle inequality
hL H −1/2 (Γ ) ≤ w − whL H −1/2 (Γ ) + whL − w

w − w hL H −1/2 (Γ )
and by using the error estimate (14.3).

Remark 14.23. The error estimate of Theorem 14.22 is not optimal with re-
1
spect to the regularity of the solution w ∈ Hpw (Γ ). Since Lemma 14.16 is
1 1 1
only valid for s ∈ (− 2 , 2 ) the higher regularity w ∈ Hpw (Γ ) is not recognized
in the error estimate. Formally, this yields the error estimate
2 3
3/2
w − whL H −1/2 (Γ ) ≤ c hL + hκL wHpw 1 (Γ ) .
When summarizing the results of Lemma 14.18 on the numerical amount

of work and the error estimate of Remark 14.23 we have to notice that it is not
possible to choose the compression parameter κ ≥ 1 in (14.39) in an optimal
way. In particular for κ = 32 we obtain in Remark 14.23 the same asymptotic
accuracy as in the error estimate (14.3) of the standard Galerkin boundary
element method, but the number of nonzero elements of the stiffness matrix
V L is O(NL (log2 N )2 ) and therefore not optimal. On the other hand, by
3/2
choosing κ = 1 we would obtain O(NL (log2 N )2 ) nonzero elements, but for a

1
sufficient regular solution w ∈ Hpw (Γ ) we will lose accuracy. The theoretical
background of this behavior is given in the proof of Lemma 14.17. There, the
moment condition (14.36) is used for piecewise constant wavelets, i.e. they are
orthogonal on constant functions. To obtain a higher order of approximation
we therefore have to require higher order moment conditions, e.g. orthogonal-
ity with respect to linear functions. This can be ensured when using piecewise
linear wavelets [76] but their construction is a quite challenging task.
14.4 Exercises
14.1 Consider the finite element stiffness matrix of Exercise 11.1 for h = 1/9,
⎛ ⎞
2 −1
⎜ −1 2 −1 ⎟
⎜ ⎟
⎜ −1 2 −1 ⎟
⎜ ⎟
⎜ −1 2 −1 ⎟
Kh = 9 ⎜⎜
⎟.
⎟
⎜ −1 2 −1 ⎟
⎜ −1 2 −1 ⎟
⎜ ⎟
⎝ −1 2 −1 ⎠
−1 2
Write Kh as a hierarchical matrix and compute the inverse Kh−1 as a hierar-

chical matrix.
14.2. The solution of the Dirichlet boundary value problem
−u (x) = f (x) for x ∈ (0, 1), u(0) = u(1) = 0
is given by
1
u(x) = (N f )(x) = G(x, y)f (y)dy for x ∈ (0, 1)
0
where G(x, y) is the associated Green function, cf. Exercise 5.2. Discuss the
Galerkin discretization of N f when using piecewise linear continuous basis
functions with respect to a uniform decomposition of (0, 1).
15
Domain Decomposition Methods
Domain decomposition methods are a modern numerical tool to handle par-

tial differential equations with jumping coefficients, and to couple different
discretization methods such as finite and boundary element methods [43].
Moreover, domain decomposition methods allow the derivation and paral-
lelization of efficient solution strategies [95] in a natural setting. For a more
detailed study of domain decomposition methods we refer, for example, to
[18, 68, 114, 139, 149].
As a model problem we consider the potential equation
−div [α(x)∇u(x)] = 0 for x ∈ Ω,

(15.1)
γ0int u(x) = g(x) for x ∈ Γ = ∂Ω
where Ω ⊂ Rd is a bounded Lipschitz domain for which a non–overlapping

domain decomposition is given, see Fig. 15.1,
*
p
Ω = Ωi, Ωi ∩ Ωj = ∅ for i = j. (15.2)
i=1
The subdomains Ωi are assumed to be Lipschitz with boundaries Γi = ∂Ωi .

By
*p
ΓS := Γi
i=1
we denote the skeleton of the domain decomposition.

We assume that in (15.1) the coefficient α(x) is piecewise constant, i.e.
α(x) = αi for x ∈ Ωi , i = 1, . . . , p. (15.3)
For an approximate solution of the boundary value problem (15.1) we will use
a boundary element method within the subdomains Ω1 , . . . , Ωq while for the
remaining subdomains Ωq+1 , . . . , Ωp a finite element method will be applied.
368 15 Domain Decomposition Methods
Ω3 Ω4
ΓS
Ω1 Ω2
Fig. 15.1. Domain decomposition with four subdomains.
By using the results of Chapter 4 the variational formulation of the Dirich-

let boundary value problem (15.1) is to find u ∈ H 1 (Ω) with γ0int u = g such
that
α(x)∇u(x)∇v(x)dx = 0 (15.4)
Ω
is satisfied for all v ∈ H01 (Ω).

Due to the non–overlapping domain decomposition (15.2) and by using
assumption (15.3) the variational formulation (15.4) is equivalent to

p
αi ∇u(x)∇v(x)dx = 0 for all v ∈ H01 (Ω).
i=1 Ωi
The application of Green’s first formula (1.5) with respect to the subdomains
Ωi for i = 1, . . . , q ≤ p results in a variational problem to find u ∈ H 1 (Ω)
with γ0int u = g such that

q
p
αi int u(x)γ int v(x)ds +
γ1,i αi ∇u(x)∇v(x)dx = 0 (15.5)
0,i x
i=1 ∂Ωi i=q+1 Ωi
is satisfied for all v ∈ H01 (Ω).

The Cauchy data γ0,i int u and γ int u of the solution u are solutions of the
1,i
boundary integral equations (6.22) on Γi = ∂Ωi , i = 1, . . . , q, i.e.
& ' & '& '
int u 1 int u
2 I − Ki
γ0,i Vi γ0,i
int u
= 1 int u
. (15.6)
γ1,i Di 2 I + Ki γ1,i
Inserting the second equation of (15.6) into the variational formulation (15.5)
this results in the variational problem to find u ∈ H 1 (Ω) with γ0int u = g and
int u ∈ H −1/2 (Γ ) for i = 1, . . . , q such that
γ1,i i
15. Domain Decomposition Methods 369

q p
int u + ( 1 I + K )γ int u, γ int v +
αi Di γ0,i α ∇u(x)∇v(x)dx = 0
i 1,i 0,i Γi i
i=1
2 i=q+1 Ωi

1
int u − ( I + K )γ int u, τ
αi Vi γ1,i i 0,i i Γi = 0
2
is satisfied for all v ∈ H01 (Ω) andτi ∈ H −1/2 (Γi ), i = 1, . . . , q.

By introducing the bilinear form
int u, . . . , γ int u; v, τ , . . . , τ )
a(u, γ1,1 1,q 1 q

q
:= int u + ( 1 I + K )γ int u, γ int v
αi Di γ0,i i 1,i 0,i Γi
i=1
2
q
int 1 int
+ αi Vi γ1,i u − ( I + Ki )γ0,i u, τi Γi
i=1
2
p
+ αi ∇u(x)∇v(x)dx
i=q+1 Ωi
we finally obtain a variational formulation to find u ∈ H 1 (Ω) with γ0int u = g

int u ∈ H −1/2 (Γ ), i = 1, . . . , q, such that
and γ1,i i
a(u, γ int
1
u; v, τ ) = 0 (15.7)
is satisfied for all v ∈ H01 (Ω) and τi ∈ H −1/2 (Γi ), i = 1, . . . , q.
Theorem 15.1. There exists a unique solution of the variational problem

(15.7).
Proof. It is sufficient to prove all assumptions of Theorem 3.8. For this we

define
X := H01 (Ω) × H −1/2 (Γ1 ) × · · · × H −1/2 (Γq )
where the norm is given by
q 2
3
p
(u, t)2X := int 2 2
γ0,i uH 1/2 (Γi ) + ti H −1/2 (Γi ) + u2H 1 (Ωi ) .
i=1 i=q+1
The boundedness of the bilinear form a(·, ·) follows from the boundedness of
all local boundary integral operators, and from the boundedness of the local
Dirichlet forms.
For arbitrary (v, τ ) ∈ X we have

q 2 3
p
a(v, τ ; v, τ ) = int v, γ int v
αi Vi τi , τi Γi + Di γ0,i 0,i Γi + αi ∇v2L2 (Ωi )
i=1 i=q+1
q
V 2 3
≥ min αi c1,i , αi c1,i , αi
D
τi 2H −1/2 (Γi ) |γ0,i
int u|2
H 1/2 (Γi )
i=1,p
i=1
⎫

p ⎬
+ ∇v2L2 (Ωi ) .
⎭
i=q+1
Due to v ∈ H01 (Ω) we therefore conclude the X–ellipticity of the bilinear form
a(·, ·) and thus the unique solvability of the variational problem (15.7).
Let
Xh := Sh1 (Ω) × Sh0 (Γ1 ) × · · · × Sh0 (Γq ) ⊂ X
be a conforming trial space of piecewise linear basis functions to approximate
the potential u ∈ H01 (Ω) and of piecewise constant basis functions to approx-
imate the local Neumann data γ1,i int u ∈ H −1/2 (Γ ), i = 1, . . . , q. All degrees of
i
freedom of the trial space Sh1 (Ω) ⊂ H01 (Ω) are depicted in Fig. 15.2.
t
t
t
BEM
t
t
t
t
t t t t t t t t t t t t t t t
t
t
t
t BEM
t
t
t
Fig. 15.2. Degrees of freedom of the trial space Sh1 (Ω) ⊂ H01 (Ω).
The global trial space Sh1 (Ω) ⊂ H01 (Ω) is decomposed into local trial spaces
Sh1 (Ωi ) := Sh1 (Ω)|Ωi ∩ H01 (Ωi ) = span{ϕ1i,k }M

k=1 ,
i
i = q + 1, . . . , p,
and into a global one
Sh1 (ΓS ) = span{ϕ1S,k }M

k=1 ,
S
which is defined with respect to the skeleton ΓS . All global degrees of freedom
are characterized in Fig. 15.2 by • while all local degrees of freedom correspond
to . From the decomposition
*
p
Sh1 (Ω) = Sh1 (ΓS ) ∪ Sh1 (Ωi )
i=q+1
it follows that a function uh ∈ Sh1 (Ω) ∩ H01 (Ω) allows the representation

MS
p
Mi
uh (x) = uS,k ϕ1S,k (x) + ui,k ϕ1i,k (x).
k=1 i=q+1 k=1
Accordingly, the coefficient vector u ∈ RM can be written as

⎛ ⎞
uS
⎜ uq+1 ⎟
⎜ ⎟ uS
u = ⎜ . ⎟ = .
⎝ .. ⎠ uL
up
Finally we introduce the trial space
Sh0 (Γi ) = span{ϕ0i,k }N

k=1 ⊂ H
i −1/2
(Γi ), i = 1, . . . , q,
of piecewise constant basis functions.

Let ug ∈ H 1 (Ω) be a bounded extension of the given Dirichlet datum
g ∈ H 1/2 (Γ ). Then the Galerkin variational formulation of (15.7) is to find
u0,h ∈ Sh1 (Ω) ∩ H01 (Ω) and ti,h ∈ Sh0 (Γi ), i = 1, . . . , q, such that
a(u0,h + ug , th ; vh , τ h ) = 0 (15.8)
is satisfied for all vh ∈ Sh1 (Ω) ∩ H01 (Ω) and τi,h ∈ Sh0 (Γi ), i = 1, . . . , q.
By applying Theorem 8.1 (Cea’s Lemma) the Galerkin variational formu-
lation (15.8) has a unique solution which satisfies the a priori error estimate
(u0 − u0,h , γ int

1
u − th )X ≤ c inf (u0 − vh , γ int
1
u − τ h )X .
(vh ,τ h )∈Xh
Hence, convergence for h → 0 will follow from the approximation properties

of the trial spaces Sh1 (Ω) and Sh0 (Γi ).
The Galerkin variational formulation (15.8) is equivalent to an algebraic
system of linear equations,
⎛ ⎞⎛ ⎞ ⎛ ⎞
Vh − 12 Mh − Kh t fB
⎜1 ⎟⎜ ⎟ ⎜ ⎟
⎝ 2 Mh + Kh Dh + ASS ALS ⎠ ⎝ uS ⎠ = ⎝ f S ⎠ (15.9)
ASL ALL uL f L
where the global stiffness matrices are given by

q
Dh [, k] = int ϕ1 , γ int ϕ1 ,
αi Di γ0,i S,k 0,i S, Γi
i=1
p
ASS [, k] = αi ∇ϕ1S,k (x)∇1S, (x)dx
i=q+1 Ωi
for k, = 1, . . . , MS , and where the local stiffness matrices are
Vh = diag Vh,i , ALL = diag Ah,i (15.10)
with
Vh,i = αi Vi ϕ0i,k , ϕ0i, Γi for k, = 1, . . . , Ni , i = 1, . . . , q,
and

Ah,i = αi ∇ϕ1i,k (x)∇ϕ1i, (x)dx for k, = 1, . . . , Mi , i = q + 1, . . . , p.
Ωi
In addition, for k = 1, . . . , MS and i = 1, . . . , q we have the block matrices

int ϕ1 , ϕ0 ,
Mh,i [, k] = αi γ0,i S,k i, Γi
int ϕ1 , ϕ0
Kh,i [, k] = αi Ki γ0,i S,k i, Γi
for = 1, . . . , N1 , while for i = q + 1, . . . , p

ASL,i [, k] = αi ∇ϕ1S,k (x)∇ϕ1i, (x)dx
Ωi
for = 1, . . . , Mi and ALS = A

SL .
The global stiffness matrix of the linear system (15.9) results from an
assembling of local stiffness matrices, which stem either from a local boundary
element or from a local finite element discretization. The vector of the right
hand side in (15.9) correspondingly results from an evaluation of a(ug , 0; ·, ·),
1 int u , ψ , = 1, . . . , N , i = 1, . . . , q,
fB,i, = αi ( + Ki )γ0,i g i, Γi i
2
q
fS, = − int u , γ int ϕ1
αi Di γ0,i g 0,i S, Γi
i=1

p
− αi ∇ug (x)∇ϕ1S, (x)dx, = 1, . . . , MS ,
i=q+1 Ωi

fL,i, = −αi ∇ug (x)∇ϕ1i, (x)dx, = 1, . . . , Mi , i = q + 1, . . . , p.
Ωi
The linear system (15.9) corresponds to the general system (13.20), hence we
can apply all iterative methods of Chapter 13.3 to solve (15.9). In particular,
when eliminating the local degrees of freedom t and uL we obtain the Schur
complement system

1 1
Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ) + ASS − ALS A−1 LL SL uS = f
A
2 2
(15.11)
where the modified right hand side is given by
1
f := f S − ( Mh + Kh )Vh−1 f B − ALS A−1
LL f L .
2
Due to (15.10) the inversion of the local stiffness matrices Vh and ALL can
be done in parallel. This corresponds to the solution of local Dirichlet bound-
ary value problems. In general we have to use local preconditioners for the
local stiffness matrices Vh,i and Ah,i . For the solution of the global Schur
complement system (15.11) where the system matrix Sh is symmetric and
positive definite, we can use a preconditioned conjugate gradient scheme. The
definition of an appropriate preconditioning matrix is then based on spectral
equivalence inequalities of the corresponding Schur complement matrices,
1 1
ShBEM := Dh + ( Mh + Kh )Vh−1 ( Mh + Kh ),
2 2
and
ShFEM := ASS − ASL A−1
LL ALS ,
and with the Galerkin discretization Dh of the hypersingular boundary inte-

gral operator, see, e.g., [36, 139].
References
1. Adams, R. A.: Sobolev Spaces. Academic Press, New York, London, 1975.
2. Ainsworth, M., Oden, J. T.: A Posteriori Error Estimation in Finite Element
Analysis. John Wiley & Sons, New York, 2000.
3. Atkinson, K. E.: The Numerical Solution of Integral Equations of the Second
Kind. Cambridge University Press, 1997.
4. Axelsson, O.: Iterative Solution Methods. Cambridge University Press,
Cambridge, 1994.
5. Axelsson, O., Barker, V. A.: Finite Element Solution of Boundary Value Prob-
lems: Theory and Computation. Academic Press, Orlando, 1984.
6. Aziz, A., Babus̆ka, I.: On the angle condition in the finite element method.
SIAM J. Numer. Anal. 13 (1976) 214–226.
7. Babus̆ka, I.: The finite element method with Lagrangian multipliers. Numer.
Math. 20 (1973) 179–192.
8. Babuška, I., Strouboulis, T.: The Finite Element Method and its Reliability.
The Clarendon Press, New York, 2001.
9. Banerjee, P. K.: The Boundary Element Methods in Engineering. McGraw,
London, 1994.
10. Bangerth, W., Rannacher, R.: Adaptive Finite Element Methods for Differential
Equations. Birkhäuser, Basel, 2003.
11. Barrett, R. et al.: Templates for the Solution of Linear Systems: Building Blocks
for Iterative Methods. SIAM, Philadelphia, 1993.
12. Bebendorf, M.: Effiziente numerische Lösung von Randintegralgleichungen unter
Verwendung von Niedrigrang–Matrizen. Dissertation, Universität des Saarlan-
des, Saarbrücken, 2000.
13. Bebendorf, M.: Approximation of boundary element matrices. Numer. Math. 86
(2000) 565–589.
14. Bebendorf, M., Rjasanow, S.: Adaptive low–rank approximation of collocation
matrices. Computing 70 (2003) 1–24.
15. Beer, G.: Programming the Boundary Element Method. John Wiley & Sons,
Chichester, 2001.
16. Bergh, J., Löfström, J.: Interpolation Spaces. An Introduction. Springer, Berlin,
New York, 1976.
17. Bespalov, A., Heuer, N.: The p–version of the boundary element method for
hypersingular operators on piecewise plane open surfaces. Numer. Math. 100
(2005) 185–209.
376 References
18. Bjørstad, P., Gropp, W., Smith, B.: Domain Decomposition. Parallel Multilevel
Methods for Elliptic Partial Differential Equations. Cambridge University Press,
1996.
19. Bonnet, M.: Boundary Integral Equation Methods for Solids and Fluids. John
Wiley & Sons, Chichester, 1999.
20. Bowman, F.: Introduction to Bessel Functions. Dover, New York, 1958.
21. Braess, D.: Finite Elemente. Springer, Berlin, 1991.
22. Braess, D.: Finite Elements: Theory, Fast Solvers and Applications in Solid
Mechanics. Cambridge University Press, 1997.
23. Brakhage, H., Werner, P.: Über das Dirichletsche Aussenraumproblem für die
Helmholtzsche Schwingungsgleichung. Arch. Math. 16 (1965) 325–329.
24. Bramble, J. H.: The Lagrange multiplier method for Dirichlet’s problem. Math.
Comp. 37 (1981) 1–11.
25. Bramble, J. H.: Multigrid Methods. Pitman Research Notes in Mathematics
Series, vol. 294, Longman, Harlow, 1993.
26. Bramble, J. H., Pasciak, J. E.: A preconditioning technique for indefinite sys-
tems resulting from mixed approximations of elliptic problems. Math. Comp.
50 (1988) 1–17.
27. Bramble, J. H., Pasciak, J. E., Steinbach, O.: On the stability of the L2 projec-
tion in H 1 (Ω). Math. Comp. 71 (2002) 147–156.
28. Bramble, J. H., Pasciak, J. E., Xu, J.: Parallel multilevel preconditioners. Math.
Comp. 55 (1990) 1–22.
29. Bramble, J. H., Zlamal, M.: Triangular elements in the finite element method.
Math. Comp. 24 (1970) 809–820.
30. Brebbia, C. A., Telles, J. C. F., Wrobel, L. C.: Boundary Element Techniques:
Theory and Applications in Engineering. Springer, Berlin, 1984.
31. Brenner, S., Scott, R. L.: The Mathematical Theory of Finite Element Methods.
Springer, New York, 1994.
32. Breuer, J.: Wavelet–Approximation der symmetrischen Variationsformulierung
von Randintegralgleichungen. Diplomarbeit, Mathematisches Institut A, Uni-
versität Stuttgart, 2001.
33. Buffa, A., Hiptmair, R.: Regularized combined field integral equations. Numer.
Math. 100 (2005) 1–19.
34. Buffa, A., Sauter, S.: Stabilisation of the acoustic single layer potential on non–
smooth domains. SIAM J. Sci. Comput. 28 (2003) 1974–1999.
35. Carstensen, C.: A unifying theory of a posteriori finite element error control.
Numer. Math. 100 (2005) 617–637.
36. Carstensen, C., Kuhn, M., Langer, U.: Fast parallel solvers for symmetric bound-
ary element domain decomposition methods. Numer. Math. 79 (1998) 321–347.
37. Carstensen, C., Stephan, E. P.: A posteriori error estimates for boundary ele-
ment methods. Math. Comp. 64 (1995) 483–500.
38. Carstensen, C., Stephan, E. P.: Adaptive boundary element methods for some
first kind integral equations. SIAM J. Numer. Anal. 33 (1996) 2166–2183.
39. Chen, G., Zhou, J.: Boundary Element Methods. Academic Press, New York,
1992.
40. Cheng, A. H.–D., Cheng, D. T.: Heritage and early history of the boundary
element method. Engrg. Anal. Boundary Elements 29 (2005) 268–302.
41. Ciarlet, P. G.: The Finite Element Method for Elliptic Problems. North–
Holland, 1978.
References 377
42. Clement, P.: Approximation by finite element functions using local regulariza-
tion. RAIRO Anal. Numer. R–2 (1975) 77–84.
43. Costabel, M.: Symmetric methods for the coupling of finite elements and bound-
ary elements. In: Boundary Elements IX (C. A. Brebbia, G. Kuhn, W. L. Wend-
land eds.), Springer, Berlin, pp. 411–420, 1987.
44. Costabel, M.: Boundary integral operators on Lipschitz domains: Elementary
results. SIAM J. Math. Anal. 19 (1988) 613–626.
45. Costabel, M., Stephan, E. P.: Boundary integral equations for mixed boundary
value problems in polygonal domains and Galerkin approximations. In: Math-
ematical Models and Methods in Mechanics. Banach Centre Publ. 15, PWN,
Warschau, pp. 175–251, 1985.
46. Dahmen, W., Prössdorf, S., Schneider, R.: Wavelet approximation methods for
pseudodifferential equations I: Stability and convergence. Math. Z. 215 (1994)
583–620.
47. Dahmen, W., Prössdorf, S., Schneider, R.: Wavelet approximation methods
for pseudodifferential equations II: Matrix compression and fast solution. Adv.
Comput. Math. 1 (1993) 259–335.
48. Darve, E.: The Fast Multipole Method: Numerical implementation. J. Comp.
Phys. 160 (2000) 195–240.
49. Dauge, M.: Elliptic boundary value problems on corner domains. Smoothness
and asymptotics of solutions. Lecture Notes in Mathematics, vol. 1341, Springer,
Berlin, 1988.
50. Dautray, R., Lions, J. L.: Mathematical Analysis and Numerical Methods for
Science and Technology. Volume 4: Integral Equations and Numerical Methods.
Springer, Berlin, 1990.
51. Demkowicz, L.: Computing with hp Adaptive Finite Elements. Vol. 1. Chapman
& Hall/CRC, Boca Raton, 2007.
52. Douglas, C. C., Haase, G., Langer, U.: A Tutorial on Elliptic PDE Solvers and
their Parallelization. SIAM, Philadelphia, 2003.
53. Duvaut, G., Lions, J. L.: Inequalities in Mechanics and Physics. Springer, Berlin,
1976.
54. Engleder, S., Steinbach, O.: Modified Boundary Integral Formulations for the
Helmholtz Equation. J. Math. Anal. Appl. 331 (2007) 396–407.
55. Erichsen, S., Sauter, S. A.: Efficient automatic quadrature in 3 − d Galerkin
BEM. Comp. Meth. Appl. Mech. Eng. 157 (1998) 215–224.
56. Faermann, B.: Local a posteriori error indicators for the Galerkin discretization
of boundary integral equations. Numer. Math. 79 (1998) 43–76.
57. Fix, G. J., Strang, G.: An Analysis of the Finite Element Method. Prentice Hall
Inc., Englewood Cliffs, 1973.
58. Fortin, M.: An analysis of the convergence of mixed finite element methods.
R.A.I.R.O. Anal. Numer. 11 (1977) 341–354.
59. Fox, L., Huskey, H. D., Wilkinson, J. H.: Notes on the solution of algebraic
linear simultaneous equations. Quart. J. Mech. Appl. Math. 1 (1948) 149–173.
60. Gatica, G. N., Hsiao, G. C.: Boundary–Field Equation Methods for a Class of
Nonlinear Problems. Pitman Research Notes in Mathematics Series, vol. 331.
Longman, Harlow, 1995.
61. Gaul, L., Kögl, M., Wagner, M.: Boundary Element Methods for Engineers and
Scientists. Springer, Berlin, 2003.
378 References
62. Giebermann, K.: Schnelle Summationsverfahren zur numerischen Lösung von In-
tegralgleichungen für Streuprobleme im R3 . Dissertation, Universität Karlsruhe,
1997.
63. Gradshteyn, I. S., Ryzhik, I. M.: Table of Integrals, Series, and Products. Aca-
demic Press, New York, 1980.
64. Greengard, L.: The Rapid Evaluation of Potential Fields in Particle Systems.
The MIT Press, Cambridge, MA, 1987.
65. Greengard, L., Rokhlin, V.: A fast algorithm for particle simulations. J. Comput.
Phys. 73 (1987) 325–348.
66. Grisvard, P.: Elliptic Problems in Nonsmooth Domains. Pitman, Boston, 1985.
67. Guiggiani, G., Gigante, A.: A general algorithm for multidimensional Cauchy
principal value integrals in the boundary element method. ASME J. Appl. Mech.
57 (1990) 906–915.
68. Haase, G.: Parallelisierung numerischer Algorithmen für partielle Differential-
gleichungen. B. G. Teubner, Stuttgart, Leipzig, 1999.
69. Hackbusch, W.: Multi–Grid Methods and Applications. Springer, Berlin, 1985.
70. Hackbusch, W.: Iterative Lösung grosser schwachbesetzter Gleichungssysteme.
B. G. Teubner, Stuttgart, 1993.
71. Hackbusch, W.: Theorie und Numerik elliptischer Differentialgleichungen.
B. G. Teubner, Stuttgart, 1996.
72. Hackbusch, W.: A sparse matrix arithmetic based on H–matrices. I. Introduc-
tion to H–matrices. Computing 62 (1999) 89–108.
73. Hackbusch, W., Nowak, Z. P.: On the fast matrix multiplication in the boundary
element method by panel clustering. Numer. Math. 54, 463–491 (1989).
74. Hackbusch, W., Wittum, G. (eds.): Boundary Elements: Implementation and
Analysis of Advanced Algorithms. Notes on Numerical Fluid Mechanics 54,
Vieweg, Braunschweig, 1996.
75. Han, H.: The boundary integro–differential equations of three–dimensional Neu-
mann problem in linear elasticity. Numer. Math. 68 (1994) 269–281.
76. Harbrecht, H.: Wavelet Galerkin schemes for the boundary element method in
three dimensions. Doctoral Thesis, TU Chemnitz, 2001.
77. Hartmann, F.: Introduction to Boundary Elements. Springer, Berlin, 1989.
78. Hestenes, M., Stiefel, E.: Methods of conjugate gradients for solving linear sys-
tems. J. Res. Nat. Bur. Stand 49 (1952) 409–436.
79. Hörmander, L.: The Analysis of Linear Partial Differential Operators I, Springer,
Berlin, 1983.
80. Hsiao, G. C., Stephan, E. P., Wendland, W. L.: On the integral equation method
for the plane mixed boundary value problem of the Laplacian. Math. Meth.
Appl. Sci. 1 (1979) 265–321.
81. Hsiao, G. C., Wendland, W. L.: A finite element method for some integral
equations of the first kind. J. Math. Anal. Appl. 58 (1977) 449–481.
82. Hsiao, G. C., Wendland, W. L.: The Aubin–Nitsche lemma for integral equa-
tions. J. Int. Equat. 3 (1981) 299–315.
83. Hsiao, G. C., Wendland, W. L.: Integral Equation Methods for Boundary Value
Problems. Springer, Heidelberg, to appear.
84. Jaswon, M. A., Symm, G. T.: Integral Equation Methods in Potential Theory
and Elastostatics. Academic Press, London, 1977.
85. Jung, M., Langer, U.: Methode der finiten Elemente für Ingenieure. B. G.
Teubner, Stuttgart, Leipzig, Wiesbaden, 2001.
References 379
86. Jung, M., Steinbach, O.: A finite element–boundary element algorithm for in-
homogeneous boundary value problems. Computing 68 (2002) 1–17.
87. Kieser, R., Schwab, C., Wendland, W. L.: Numerical evaluation of singular and
finite–part integrals on curved surfaces using symbolic manipulation. Computing
49 (1992) 279–301.
88. Kress, R.: Linear Integral Equations. Springer, Heidelberg, 1999.
89. Kupradze, V. D.: Three–dimensional problems of the mathematical theory of
elasticity and thermoelasticity. North–Holland, Amsterdam, 1979.
90. Kythe, P. K.: Fundamental Solutions for Differential Operators and Applica-
tions. Birkhäuser, Boston, 1996.
91. Ladyzenskaja, O. A.: Funktionalanalytische Untersuchungen der Navier–
Stokesschen Gleichungen. Akademie–Verlag, Berlin, 1965.
92. Ladyzenskaja, O. A., Ural’ceva, N. N.: Linear and quasilinear elliptic equations.
Academic Press, New York, 1968.
93. Lage, C., Sauter, S. A.: Transformation of hypersingular integrals and black–box
cubature. Math. Comp. 70 (2001) 223–250.
94. Lage, C., Schwab, C.: Wavelet Galerkin algorithms for boundary integral equa-
tions. SIAM J. Sci. Comput. 20 (1999) 2195–2222.
95. Langer, U.: Parallel iterative solution of symmetric coupled fe/be equations via
domain decomposition. Contemp. Math. 157 (1994) 335–344.
96. Langer, U., Pusch, D.: Data–sparse algebraic multigrid methods for large scale
boundary element equations. Appl. Numer. Math. 54 (2005) 406–424.
97. Langer, U., Steinbach, O.: Coupled boundary and finite element tearing and
interconnecting methods. In: Domain Decomposition Methods in Science and
Engineering (R. Kornhuber et. al. eds.), Lecture Notes in Computational Science
and Engineering, vol. 40, Springer, Heidelberg, pp. 83–97, 2004.
98. Langer, U., Steinbach, O., Wendland, W. L.: Computing and Visualization in
Science. Special Issue on Fast Boundary Element Methods in Industrial Appli-
cations. Volume 8, Numbers 3–4, 2005.
99. Maischak, M.: hp–Methoden für Randintegralgleichungen bei 3D–Problemen.
Theorie und Implementierung. Doctoral Thesis, Universität Hannover, 1995.
100. Maischak, M., Stephan, E. P.: The hp–version of the boundary element method
in R3 . The basic approximation results. Math. Meth. Appl. Sci. 20 (1997) 461–
476.
101. Maue, A. W.: Zur Formulierung eines allgemeinen Beugungsproblems durch
eine Integralgleichung. Z. f. Physik 126 (1949) 601–618.
102. Mazya, V. G.: Boundary integral equations. In: Analysis IV (V. G. Mazya,
S. M. Nikolskii eds.), Encyclopaedia of Mathematical Sciences, vol. 27, Springer,
Heidelberg, pp. 127–233, 1991.
103. McLean, W.: Strongly Elliptic Systems and Boundary Integral Equations.
Cambridge University Press, 2000.
104. McLean, W., Steinbach, O.: Boundary element preconditioners for a hyper-
singular boundary integral equation on an intervall. Adv. Comput. Math. 11
(1999) 271–286.
105. McLean, W., Tran, T.: A preconditioning strategy for boundary element
Galerkin methods. Numer. Meth. Part. Diff. Eq. 13 (1997) 283–301.
106. Nec̆as, J.: Les Methodes Directes en Theorie des Equations Elliptiques. Masson,
Paris und Academia, Prag, 1967.
107. Nedelec, J. C.: Integral equations with non integrable kernels. Int. Eq. Operator
Th. 5 (1982) 562–572.
380 References
108. Of, G., Steinbach, O.: A fast multipole boundary element method for a modified
hypersingular boundary integral equation. In: Analysis and Simulation of Mul-
tifield Problems (W. L. Wendland, M. Efendiev eds.), Lecture Notes in Applied
and Computational Mechanics 12, Springer, Heidelberg, 2003, pp. 163–169.
109. Of, G., Steinbach, O., Wendland, W. L.: Applications of a fast multipole
Galerkin boundary element method in linear elastostatics. Comput. Vis. Sci.
8 (2005) 201–209.
110. Of, G., Steinbach, O., Wendland, W. L.: The fast multipole method for the
symmetric boundary integral formulation. IMA J. Numer. Anal. 26 (2006) 272–
296.
111. Petersdorff, T. von, Stephan, E. P.: On the convergence of the multigrid method
for a hypersingular integral equation of the first kind. Numer. Math. 57 (1990)
379–391.
112. Plemelj, J.: Potentialtheoretische Untersuchungen. Teubner, Leipzig, 1911.
113. Prössdorf, S., Silbermann, B.: Numerical Analysis for Integral and Related
Operator Equations. Birkhäuser, Basel, 1991.
114. Quarteroni, A., Valli, A.: Domain Decomposition Methods for Partial Differ-
ential Equations. Oxford Science Publications, 1999.
115. Rathsfeld, A.: Quadrature methods for 2D and 3D problems. J. Comput. Appl.
Math. 125 (2000) 439–460.
116. Reidinger, B., Steinbach, O.: A symmetric boundary element method for the
Stokes problem in multiple connected domains. Math. Meth. Appl. Sci. 26 (2003)
77–93.
117. Rjasanow, S., Steinbach, O.: The Fast Solution of Boundary Integral Equa-
tions. Mathematical and Analytical Techniques with Applications to Engineer-
ing. Springer, New York, 2007.
118. Rudin, W.: Functional Analysis. McGraw–Hill, New York, 1973.
119. Ruotsalainen, K., Wendland, W. L.: On the boundary element method for some
nonlinear boundary value problems. Numer. Math. 53 (1988) 299–314.
120. Saad, Y., Schultz, M. H.: A generalized minimal residual algorithm for solving
nonsymmetric linear systems. SIAM J. Sci. Stat. Comput. 7 (1985) 856–869.
121. Sauer, T., Xu, Y.: On multivariate Lagrange interpolation. Math. Comp. 64
(1995) 1147–1170.
122. Sauter, S. A.: Variable order panel clustering. Computing 64 (2000) 223–261.
123. Sauter, S. A., Schwab, C.: Quadrature of hp–Galerkin BEM in R3 . Numer.
Math. 78 (1997) 211–258.
124. Sauter, S. A., Schwab, C.: Randelementmethoden. Analyse, Numerik und
Implementierung schneller Algorithmen. B. G. Teubner, Stuttgart, Leipzig,
Wiesbaden, 2004.
125. Schanz, M., Steinbach, O. (eds.): Boundary Element Analysis: Mathematical
Aspects and Applications. Lecture Notes in Applied and Computational Me-
chanics, vol. 29, Springer, Heidelberg, 2007.
126. Schatz, A. H., Thomée, V., Wendland, W. L.: Mathematical Theory of Finite
and Boundary Element Methods. Birkhäuser, Basel, 1990.
127. Schneider, R.: Multiskalen– und Wavelet–Matrixkompression: Analysisbasierte
Methoden zur effizienten Lösung grosser vollbesetzter Gleichungssysteme. Ad-
vances in Numerical Mathematics. B. G. Teubner, Stuttgart, 1998.
128. Schulz, H., Steinbach, O.: A new a posteriori error estimator in adaptive direct
boundary element methods. Calcolo 37 (2000) 79–96.
References 381
129. Schulz, H., Wendland, W. L.: Local a posteriori error estimates for boundary
element methods. In: ENUMATH 2007, World Sci. Publ., River edge, pp. 564–
571, 1998.
130. Schwab, C.: p– and hp–Finite Element Methods. Theory and Applications in
Solid and Fluid Mechanics. Clarendon Press, Oxford, 1998.
131. Schwab, C., Suri, M.: The optimal p–version approximation of singularities on
polyhedra in the boundary element method. SIAM J. Numer. Anal. 33 (1996)
729–759.
132. Schwab, C., Wendland, W. L.: On numerical cubatures of singular surface
integrals in boundary element methods. Numer. Math. 62 (1992) 343–369.
133. Scott, L. R., Zhang, S.: Finite element interpolation of nonsmooth functions
satisfying boundary conditions. Math. Comp. 54 (1990) 483–493.
134. Sirtori, S.: General stress analysis method by means of integral equations and
boundary elements. Meccanica 14 (1979) 210–218.
135. Sloan, I. H.: Error analysis of boundary integral methods. Acta Numerica 92
(1992) 287–339.
136. Steinbach, O.: Fast evaluation of Newton potentials in boundary element meth-
ods. East–West J. Numer. Math. 7 (1999) 211–222.
137. Steinbach, O.: On the stability of the L2 projection in fractional Sobolev spaces.
Numer. Math. 88 (2001) 367–379.
138. Steinbach, O.: On a generalized L2 projection and some related stability esti-
mates in Sobolev spaces. Numer. Math. 90 (2002) 775–786.
139. Steinbach, O.: Stability estimates for hybrid coupled domain decomposition
methods. Lecture Notes in Mathematics 1809, Springer, Heidelberg, 2003.
140. Steinbach, O.: Numerische Näherungsverfahren für elliptische Randwertprob-
leme. Finite Elemente und Randelemente. B. G. Teubner, Stuttgart, Leipzig,
Wiesbaden, 2003.
141. Steinbach, O.: A robust boundary element method for nearly incompressible
elasticity. Numer. Math. 95 (2003) 553–562.
142. Steinbach, O.: A note on the ellipticity of the single layer potential in two–
dimensional linear elastostatics. J. Math. Anal. Appl. 294 (2004) 1–6.
143. Steinbach, O.: Lösungsverfahren für lineare Gleichungssysteme. Algorithmen
und Anwendungen. B. G. Teubner, Stuttgart, Leipzig, Wiesbaden, 2005.
144. Steinbach, O., Wendland, W L.: The construction of some efficient precondi-
tioners in the boundary element method. Adv. Comput. Math. 9 (1998) 191–216.
145. Steinbach, O., Wendland, W. L.: On C. Neumann’s method for second order
elliptic systems in domains with non–smooth boundaries. J. Math. Anal. Appl.
262 (2001) 733–748.
146. Stephan, E. P.: The h–p version of the boundary element method for solving
2– and 3–dimensional problems. Comp. Meth. Appl. Mech. Eng. 133 (1996)
183–208.
147. Stephan, E. P.: Multilevel methods for the h–, p–, and hp–versions of the
boundary element method. J. Comput. Appl. Math. 125 (2000) 503–519.
148. Stroud, A. H.: Approximate Calculations of Multiple Integrals. Prentice Hall,
Englewood Cliffs, 1973.
149. Toselli, A., Widlund, O.: Domain Decomposition Methods–Algorithms and
Theory. Springer, Berlin, 2005.
150. Triebel, H.: Höhere Analysis. Verlag Harri Deutsch, Frankfurt/M., 1980.
151. Tyrtyshnikov, E. E.: Mosaic–skeleton approximations. Calcolo 33 (1996) 47–57.
382 References
152. Verchota, G.: Layer potentials and regularity for the Dirichlet problem for
Laplace’s equation in Lipschitz domains. J. Funct. Anal. 59 (1984) 572–611.
153. Verfürth, R.: A Review of A Posteriori Error Estimation and Adaptive Mesh–
Refinement. John Wiley & Sons, Chichester, 1996.
154. Vladimirov, V. S.: Equations of Mathematical Physics. Marcel Dekker, New
York, 1971.
155. van der Vorst, H. A.: Bi–CGSTAB: A fast and smoothly converging variant
of Bi–CG for the solution of nonsymmetric linear systems. SIAM J. Sci. Stat.
Comput. 13 (1992) 631–644.
156. Walter, W.: Einführung in die Theorie der Distributionen. BI Wissenschaftsver-
lag, Mannheim, 1994.
157. Wendland, W. L.: Elliptic Systems in the Plane. Pitman, London, 1979.
158. Wendland, W. L. (ed.): Boundary Element Topics. Springer, Heidelberg, 1997.
159. Wendland, W. L., Zhu, J.: The boundary element method for three–
dimensional Stokes flow exterior to an open surface. Mathematical and Com-
puter Modelling 15 (1991) 19–42.
160. Wloka, J.: Funktionalanalysis und Anwendungen. Walter de Gruyter, Berlin,
1971.
161. Wloka, J.: Partielle Differentialgleichungen. B. G. Teubner, Stuttgart, 1982.
162. Xu, J.: An introduction to multilevel methods. In: Wavelets, multilevel methods
and elliptic PDEs. Numer. Math. Sci. Comput., Oxford University Press, New
York, 1997, pp. 213–302.
163. Yosida, K.: Functional Analysis. Springer, Berlin, Heidelberg, 1980.
Index
C ∞ (Ω), C0∞ (Ω), 19 Stokes system, 14

C k,κ (Ω), 20 boundary condition
H∗1 (Ω), 67 Dirichlet, 3
H s (Γ ), 35 Neumann, 3
s (Γ0 ), 37
H s (Γ0 ), H Robin, 3
s
H (Ω), H s (Ω), H0s (Ω), 33 sliding, 15
s boundary element, 229
Hpw (Γ ), 37
−1/2 boundary stress operator, 7
H∗ (Γ ), 139
1/2
H∗ (Γ ), 144 boundary value problem
1/2 Dirichlet, 61, 76, 172, 243, 263, 327,
H∗∗ (Γ ), 147
loc 367
L1 (Ω), 22
exterior, 181
Lp (Ω), L∞ (Ω), 21
mixed, 70, 79, 179, 281
L2,0 (Ω), 80
Neumann, 67, 77, 175, 253, 274
Wps (Ω), 24
Robin, 71, 181, 287
D(Ω), D (Ω), 29
Brakhage–Werner formulation, 185
S(Rd ), S (Rd ), 30
Bramble–Hilbert lemma, 27, 39
A–orthogonal vectors, 291
admissible Calderón projection, 137
maximally, 331 capacity, 142
Airy’s stress function, 9, 97 logarithmic, 142
approximation Cauchy data, 90
of the linear form, 190 Cauchy–Schwarz inequality, 311
operator, 191 Cea’s lemma, 189
property, 189, 220, 224, 237, 241 CG method, 295
Aubin–Nitsche trick, 244, 266, 276 Bramble/Pasciak, 325
Clement operator, 226
BBL condition, 196 cluster
Bessel potential, 32 admissible, 331
Betti’s first formula, 5 tree, 330
bilinear form coercive operator, 57
linear elasticity, 6 condition number
potential equation, 2 BEM stiffness matrix, 269
384 Index
FEM stiffness matrix, 250 Galerkin–Bubnov method, 187

conjugate gradients, 295 Galerkin–Petrov method, 193
conjugate vectors, 291 globally quasi–uniform, 205
conormal derivative Gram–Schmidt orthogonalization, 293
linear elasticity, 6 Green’s formula
potential equation, 2 first, 2
Stokes system, 14 second, 3
contraction, 149 Gårdings inequality, 57
criterion of Fortin, 194, 199
Haar wavelets, 353
decomposition, 203, 229 hierarchical matrices, 332
admissible, 204, 230 Hooke’s law, 5
globally quasi–uniform, 205, 231 hypersingular operator, 128, 163
locally quasi–uniform, 205, 231 integration by parts, 131, 134, 164
shape regular, 205
derivative incompressible materials, 12
generalized, 23 index set, 203
in the sense of distributions, 30 indirect approach, 172
diameter, 204, 231 inequality
direct approach, 171 Cauchy–Schwarz, 22, 311
Dirichlet to Neumann map, 148, 180 Hölder, 21
distribution, 29 Korn’s first, 74
domain decomposition, 367 Korn’s second, 76
double layer potential, 124 Minkowski, 21
adjoint, 120, 157 Poincaré, 27
integration by parts, 164 interpolation, 218
linear elastostatics, 162 inverse inequality
dual space, 25 global, 217, 240, 242
duality pairing, 22, 36 local, 213, 239
jump relation
edge, 203
adjoint double layer potential, 123
elliptic operator, 1, 46
double layer potential, 127
single layer potential, 120
far field, 331
finite element, 203 Korn’s inequality
form function first, 74
Bubble–, 215 second, 76
constant, 211
linear, 212 Lagrange
quadratic, 214 functional, 53
Fourier transform, 30 multiplier, 52, 64
Fredholm alternative, 58 multipliers, 255
fundamental solution, 90 Lamé constants, 6
Helmholtz operator, 106, 109 Lax–Milgram lemma, 46
Laplace operator, 96 Lipschitz domain, 20
linear elasticity, 100 load vector, 268
Stokes system, 103 locally quasi–uniform, 205
Galerkin orthogonality, 188, 193, 198 maximally admissible, 331

Index 385
mesh size Laplace, 89

global, 205, 231 linear elasticity, 97
local, 204, 230 Stokes system, 101, 105
method of conjugate gradients, 295 Riesz map, 45
preconditioned, 298 rigid body motions, 7, 11
mixed formulation, 52 rotation, 11
moment condition, 353
multi index, 19 scaling condition, 67, 80
multilevel operator, 308 Schur complement system, 196, 257,
283, 320, 373
natural density, 142, 144 Schur Lemma, 310
near field, 331 series expansion
neighboring elements, 205, 231 ACA, 344
Neumann series, 173, 177 fundamental solution, 340
Newton potential, 111 Taylor series, 337
node, 203, 230 shape regular, 205, 232
norm equivalence, 211 single layer potential, 118, 157
null space, 48 Sobolev space, 24
Sobolev–Slobodeckii norm, 24, 36
operator equation, 41 solvability condition, 4, 8, 12, 48, 50
orthogonal complement, 48 Somigliana identity, 156
orthogonal space, 48 Sommerfeld radiation condition, 16
spectral condition number, 296
partial differential equation BEM stiffness matrix, 278
Bi–Laplace, 9 stability condition, 50, 54, 64, 174
Helmholtz, 183 discrete, 193, 196, 257, 301, 305
Helmholtz operator, 15 Stokes problem, 81
Laplace operator, 3, 111 Steklov–Poincaré operator, 148, 284
linear elasticity, 5 stiffness matrix, 249, 268, 277
linear elastostatics, 156 local, 252
Stokes system, 12, 165 strain tensor, 5
partial differential operator, 1 Strang lemma, 191, 192
partition of unity, 253 stress function of Airy, 97
plain strain, 10 stress tensor, 5
plain stress, 9 support of a function, 19
Poisson ratio, 5 surface curl, 133
preconditioning, 297, 299 symmetric approximation, 284
multilevel, 306 symmetric formulation, 179, 281
with integral operators, 302
projection theorem
H 1 –, 220 closed range theorem, 48
H σ –, 237, 241 equivalence theorem of Sobolev, 26
L2 –, 219, 233, 241 imbedding theorem, 25
interpolation theorem, 34
quasi interpolation operator, 226 inverse trace theorem, 38
Lax–Milgram lemma, 46
radius, 205 of Gauss and Ostrogradski, 2
reference element, 206, 209, 231 Riesz representation theorem, 43
representation formula trace theorem, 38
386 Index
trace, 2 uniform elliptic operator, 1

trial space
variational problem, 42
linear, 216, 238, 275
volume, 204, 230
piecewise constant, 233, 263
quadratic, 222 Young modulus, 5

Numerical Approximation Methods For Elliptic Boundary Value Problems PDF

Hochgeladen von

Dokumentinformationen

Originaltitel

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Numerical Approximation Methods For Elliptic Boundary Value Problems PDF

Hochgeladen von

Copyright:

Verfügbare Formate

Numerical Approximation Methods

for Elliptic Boundary Value Problems

Finite and Boundary Elements

English version published by Springer Science+Business Media, LLC

ISBN 978-0-387-31312-2 e-ISBN 978-0-387-68805-3

Library of Congress Control Number: 2007936614

Mathematics Subject Classiﬁcation (2000): 65N30, 65N38

c 2008 Springer Science+Business Media, LLC

Printed on acid-free paper.

Lagrange multipliers. The variational formulation of boundary value problems

Graz, August 2007 Olaf Steinbach

1 Boundary Value Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

4 Variational Formulations of Boundary Value Problems . . . . . 59

4.1.5 Robin Boundary Value Problems . . . . . . . . . . . . . . . . . . . . 71

6 Boundary Integral Operators . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

7 Boundary Integral Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171

8 Approximation Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187

9 Finite Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 203

10 Boundary Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229

11 Finite Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243

12 Boundary Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263

13 Iterative Solution Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291

14 Fast Boundary Element Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 327

14.3 Wavelets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 351

15 Domain Decomposition Methods . . . . . . . . . . . . . . . . . . . . . . . . . . 367

In this chapter we describe some stationary boundary value problems with

1.1 Potential Equation

is the interior boundary trace of a given function f (x), x ∈ Ω.

When multiplying the partial diﬀerential operator (1.1) with a suﬃciently

Applying integration by parts, see (1.4), we obtain

and therefore Green’s ﬁrst formula

by using the symmetric bilinear form

as well as the interior conormal derivative

As in (1.5) we ﬁnd, by exchanging the role of u and v, the analogue Green’s

Combining this with (1.5) we therefore obtain Green’s second formula

which holds for arbitrary but suﬃciently smooth functions u and v.

Let Γ = Γ D ∪ Γ N ∪ Γ R be a disjoint decomposition of the boundary

(Lu)(x) = f (x) for x ∈ Ω, (1.10)

the Dirichlet boundary conditions

γ0int u(x) = gD (x) for x ∈ ΓD , (1.11)

the Neumann boundary conditions

γ1int u(x) = gN (x) for x ∈ ΓN , (1.12)

and the Robin boundary conditions

γ1int u(x) + κ(x)γ0int u(x) = gR (x) for x ∈ ΓR (1.13)

where f , gD , gN , gR , and κ are some given functions. The boundary value

γ1int u(x) + G(γ0int u, x) = gR (x) for x ∈ ΓR

where G(v, ·) is some given function, e.g. G(v, ·) = v 3 or G(v, ·) = v 4 .

(Lv1 )(x) = 0 for x ∈ Ω, γ1int v1 (x) = 0 for x ∈ Γ. (1.14)

Applying Green’s second formula (1.8) we then obtain the orthogonality

(Lu)(x) = f (x) for x ∈ Ω, γ1int u(x) = gN (x) for x ∈ Γ, (1.16)

Since there exists a non–trivial solution v1 (x) = 1 for x ∈ Ω of the homoge-

to be also a solution of the Neumann boundary value problem (1.16). The

1.2 Linear Elasticity

Taking the sum for i = 1, 2, 3 we obtain Betti’s ﬁrst formula

with the bilinear form

and with the conormal derivative

From this we obtain

for x ∈ Ω, i = 1, 2, 3. By introducing the Lamé constants

we ﬁnally conclude the Navier system

−µ∆u(x) − (λ + µ)grad div u(x) = f (x) for x ∈ Ω. (1.25)

The bilinear form (1.22) can be written as

implying the symmetry of the bilinear form a(·, ·).

γ0int ui (x) = gD,i (x) for x ∈ ΓD,i ,

(r, ϕ) in terms of the gradient of u(x1 , x2 ), i.e. ﬁnd

−uk (x) = λuk (x) for x ∈ (0, 1), uk (0) = uk (1) = 0.